DEV Community: Givi Pataridze

Your blog no longer has to look like a template

Givi Pataridze — Tue, 13 Oct 2026 18:04:56 +0000

Until now, every GoodSender blog started from the same template. Your posts were your own, but the page around them was shared with every other blog on the platform.

Starting today, you can describe the blog you want and get a design made for it. Open Design in your GoodSender dashboard and tell the assistant what you have in mind. It designs every page of the blog: the landing page, the archive, tag pages, and the post page. Then you refine the result in conversation.

The designer is available on every GoodSender account, with no usage limits at launch.

Start with a description, a link, or your brand files

Describe how the blog should feel: the mood, the colors, the typefaces, the people who read it. A couple of sentences is enough to start:

"A blog for a small product studio. Calm, lots of white space, one warm accent color, serif headlines. Put a subscribe form under every post."

If you already have a brand, give the assistant something to work from:

Attach your logo as a PNG, JPEG, GIF, or WebP, and it can use the image as the blog's logo or favicon.
Attach a brand guide as a PDF or a text file.
Paste a link to a public page, such as your company website. The assistant reads the text of that page. A page that only renders in the browser may come back empty.

The assistant designs the four page types together. By default, it keeps the header, footer, navigation, and typography consistent across them, so the blog reads as one design system rather than four separate pages. If you want the pages to differ more, say so.

Same posts, five designs

Each design below was made with the designer, and all five show the same six demo posts.

A terminal-style blog for developers: a compact post list, read times, and tags.

An editorial look for a consultant: a featured latest post and a weekly subscribe block.

A product updates blog for a startup: tag chips, a subscribe form up top, and cards with covers.

A loud, playful design with big type and numbered cards.

A retro desktop theme, where the post list and the subscribe form live in windows.

Refine it the way you would with a designer

The preview updates after each change. Ask for anything: colors, fonts, images, text.

Some changes are easier to point at than to describe. Press Pick, click an element in the preview, and write what should change, for example "make this smaller" or "move this under the title". The assistant edits that element on that page.

Then test the design against content it was not drawn for. Switch the preview between the landing, archive, tag, and post pages. Resize it to see how the layout behaves on a narrow screen. Open Contents to fill the preview with sample posts: long and short titles, posts with and without covers, tables, lists, and code.

Each change becomes a checkpoint. If a direction doesn't work, go back to an earlier one. Going back removes the checkpoints after it. When you like the result, save the design and make it live. You can keep several designs and change which one is live later.

A new look on the same SEO and AEO foundation

A redesign shouldn't cost you the work that makes a blog findable. The design decides how your blog looks. GoodSender still adds the tags that tell search engines and AI systems what each page is.

The design decides	GoodSender adds to every page
Layout, colors, and typography	Page title and meta description
Header, footer, navigation, and logo	Canonical URL
How post lists and cards look	Open Graph tags for social cards
The subscribe form's look and wording	Structured data (JSON-LD)

Your RSS feed, sitemap, and llms.txt are generated outside the design, so a redesign doesn't touch them. Pages are rendered on the server as HTML, so crawlers and AI assistants can read your posts without running an app first. The pages stay lightweight.

Or design it from your own agent

The same design tools are available through the GoodSender MCP server. If you already work with Claude or Codex, ask it to design your blog. It can create the design, refine it, and make it live from the same chat.

GoodSender vs Ghost vs Substack vs beehiiv: an honest comparison

Givi Pataridze — Fri, 09 Oct 2026 05:18:52 +0000

If you want to earn from your writing or grow an audience inside a platform's network, pick Ghost, Substack or beehiiv. GoodSender is built for a different job: a blog on your own domain with free, reliable email to your subscribers. You write in the editor yourself, or let Claude or Codex help. Below is what the four have in common, what GoodSender adds, and where the others go further. Everything is based on each vendor's own pricing pages and docs, checked October 8, 2026.

What all four have in common

	GoodSender	Ghost	Substack	beehiiv
Free way to start	Free on your own domain	Free if you self-host; Ghost(Pro) from $18/month	Free	Free plan, up to 2,500 subscribers
Your own domain	Yes, required to go public	Free custom domain	One time $50	Paid plans
Site and search basics	Sitemap, RSS, canonical tags, no setup	Advanced SEO tools, RSS	SEO settings	Website builder on every plan
Email to subscribers	Yes	Yes	Yes	Yes

The basics are covered everywhere: a site for your posts, your own domain, and email to subscribers. The differences are in cost, workflow and growth tools.

What GoodSender adds

	GoodSender	Ghost	Substack	beehiiv
Email pricing	Free up to 100,000 emails a month, priced on emails sent	Rises with members: $105/month at 10,000 (Publisher)	Free; 10% of paid subscriptions	Rises with subscribers: $139/month at 10,000 (Lite)
Work with Claude or Codex	Official MCP and CLI: draft, preview, publish	No official MCP server	Official API is read only	Official MCP writes drafts; you publish in the app
Share on publish	LinkedIn, X, Bluesky, Mastodon, dev.to, Hashnode	Social web distribution	Video to YouTube and LinkedIn	Not listed

Email that stays free as you grow. A weekly post to 10,000 subscribers is about 43,000 emails a month: free on GoodSender, and $10 a month once you pass 100,000 emails. Every subscriber confirms by double opt-in, inactive readers are asked to reconfirm after 120 days, and unsubscribes and complaints are suppressed automatically. Sending runs on Laneful, built by the team behind SendGrid.

AI help when you want it. Write everything yourself, or ask Claude or Codex to help with a draft, a preview or the publish step. beehiiv's MCP stops at drafts that "must be published from the app". Ghost has no official MCP server, and Substack's API only reads public data.

One publish, many places. A post can be announced on four social networks and republished on dev.to and Hashnode with a canonical link back to your blog. Pages need no search setup.

Where Ghost, Substack and beehiiv go further

	GoodSender	Ghost	Substack	beehiiv
Paid subscriptions	No	Yes, 0% Ghost fee	Yes, Substack keeps 10%	Yes, 0% take rate
Discovery network	No	Recommendations, Explore network, referrals	Recommendations, Notes, app, referrals	Recommendation network, ad network
Comments and community	No	Comments	Comments, chat, Notes	Not listed
Themes and design	GoodSender templates; AI design generation in development	Marketplace and custom themes (Publisher and up)	Not listed	Website builder
Import existing posts	Subscribers only	Not listed	Posts and subscribers from 12+ platforms	Not listed

Sources: Ghost pricing, Ghost on GitHub; Substack features, custom domain, SEO, API terms; beehiiv pricing, beehiiv MCP. "Not listed" means we did not find it on the vendor's pricing or feature pages.

Which one should you choose?

Ghost: paid memberships, deep design control or self-hosting.
Substack: readers through its network and app, at 10% of paid subscriptions.
beehiiv: growth and monetization tools such as recommendation and ad networks.
GoodSender: your own domain, free email to your readers, and AI help when you want it.

FAQ

Is there a free blog platform with email on your own domain?

GoodSender: the blog on your own domain is free, and so are up to 100,000 emails a month to your subscribers. Substack is also free to send, with a one time $50 for a custom domain.

Which blog platform works best with Claude?

Of these four, only GoodSender has an official MCP server that lets Claude help with the whole post, from draft to publish. beehiiv's official MCP creates drafts that you publish in beehiiv.

Start your blog on GoodSender if these trade-offs work for you.

Why readers stop trusting AI-polished writing

Givi Pataridze — Tue, 06 Oct 2026 08:52:52 +0000

You open a post on a topic you care about. Three lines in, something feels off. The rhythm is too even. Every paragraph closes with a neat little lesson. You keep scrolling, but you no longer trust the author.

The topic did not get worse. Your trust in the text did.

Bynder ran a test with 2,000 people in the US and UK in 2024. Readers compared two short articles, one written by ChatGPT and one by a person. Only half could tell which one the machine wrote. Without labels, 56% preferred the AI version. Yet 52% said they felt less engaged once they learned a text was AI-generated. So readers are not great detectors. What changes their attitude is the belief that nobody wrote the text.

That was 2024, and the gap has only become clearer. In a July 2026 DoubleVerify and YouGov survey of 2,000 UK consumers, 92% said AI-generated content undermines trust online. Only 41% felt confident they could spot it.

AI is the secretary, not the author

I think of AI as a good secretary. It fixes my grammar, tightens a paragraph, and points out where an idea needs more room. It does not decide what the post is about. The direction, the main thought, and anything new in it have to come from me. When that order flips, I think readers feel it, even if they cannot point to it.

Wikipedia editors have collected what they notice into a guide called Signs of AI writing, maintained by WikiProject AI Cleanup. It lists inflated claims about significance, vague attributions, forced groups of three, and the overused em dash. The guide also warns that it is not a ban on words or punctuation, and that the signs matter in combination.

Some of these signs are already aging. The em dash, the most famous one, is disappearing from new models. In a Graphite study reported by VentureBeat in September 2026, Claude Opus 5.5 used about 99% fewer em dashes than Opus 5, and far fewer than the human articles in the same test. Arena measured a 95% drop on its own prompts. A text without em dashes proves nothing anymore. Graphite still found 2,548 other words and phrases that Opus 5.5 uses at least twice as often as human writers.

Someone turned that guide into an agent skill. Humanizer is an open-source skill for Claude Code, Codex, and similar agents that rewrites text to remove those patterns while keeping the meaning. It is a good start and it cuts some filler. Still, I see three problems it does not fully solve.

Problem one: it is too long

Real people rarely write texts this long. A person with something to say usually says it and stops, while a model keeps going.

One study suggests why length gets rewarded. In 2023, Saito and colleagues found that GPT-4, used as a judge of answers, favored longer responses more strongly than human judges did. When machines grade machines, length can look like quality.

The habit has not gone away with newer models. In the same Arena comparison, average answers grew from 453 words with Opus 5 to 481 with Opus 5.5, even as sentences got shorter.

The best antidote I know is older than AI drafts. «Пиши, сокращай» ("Write, Shorten") by Maxim Ilyakhov and Lyudmila Sarycheva is a Russian-language book on clear business writing. Its first edition came out in 2016 and sold 500,000 copies. The 2025 edition was reworked and adds a chapter on the editor's career and neural networks. Its core idea fits on a sticky note: find the words that carry no meaning and remove them. That works on an AI draft as well as on a human one. The book's method is already available as an agent skill: sokrati is an open-source Claude skill with versions in seven languages, and the English one is called shorten.

Problem two: it has no mistakes

AI almost always follows the rules. A polished AI draft has every comma in place and every word spelled right. People are not like that.

When I see a typo, I know the text was not written by AI. Yes, a mistake costs a writer some credibility. Today I think that cost is worth paying. A human with a few errors is worth more to readers than a flawless text with no person behind it. What matters is the signal the reader picks up without noticing it: a human wrote this.

Research narrows this idea. In 2024, Bluvstein, Zhao, Barasch, and Schroeder ran seven studies with 3,399 participants on customer-service chats. Agents who made a typo and then corrected it were seen as more human and more helpful than agents who made no typos, and more than agents who left the typo in place. In this study, the correction did the humanizing. A typo left in place did not help.

So here is an idea for a skill. Call it a mistakaizer: it would add a few human imperfections to an AI-polished text. A typo fixed in the next line, or a paragraph that runs a little long. The goal is that the reader feels, without thinking about it, that a person is on the other side.

I am not sure it should exist. If readers trust typos because machines do not make them, planting typos on purpose spends that trust. On the other hand, if your own draft had those quirks before the AI smoothed them away, restoring them is closer to repair than to forgery.

Problem three: it is the wrong English

Many people writing in English are not native speakers, and many native speakers are not American. Their phrasing and their examples differ, and AI tends to sand those differences away.

Detectors flag non-native writers. A 2023 Stanford study in Patterns tested seven GPT detectors. More than half of 91 TOEFL essays by non-native speakers were flagged as AI-generated, while over 90% of essays by US eighth-graders were correctly classified as human. Simpler vocabulary triggered the false positives. Those were 2023 detectors. Originality.ai, one of the tools tested, disputes the study and reports a 5% false positive rate on non-native essays in its own 2025 test. Even that means one non-native essay in twenty gets flagged. This is a different problem from style, but it lands on the same people.
AI suggestions pull writing toward American style. In a Cornell study presented at CHI 2025, 118 participants from the US and India wrote about their culture. The assistant suggested pizza for favorite foods and Christmas for favorite holidays. When an Indian participant started typing the name of Shah Rukh Khan, it offered Shaquille O'Neal.
Polishing erases cultural markers. A 2026 study of Indian, Singaporean, and Nigerian English analyzed 22,350 model outputs from tasks like email polishing. Models removed about 10% of cultural markers overall, and politeness conventions were hit hardest. In this preprint, explicit instructions to preserve them cut the erasure by about 29%.

The Cornell team quoted how Indian participants described Diwali without and with the assistant:

Without AI	With AI
"worship goddess Laxmi"	"eat traditional Indian breakfast items"
"pop crackers and eat sweets"	"a time filled with happiness and warmth"

Usage is uneven, too. Kobak and colleagues studied over 15 million PubMed abstracts and estimated that at least 13.5% of 2024 abstracts were processed with LLMs. The lower bound was around 5% for the UK and Australia and around 20% for China, South Korea, and Taiwan. The authors suggest non-native speakers may adopt LLM help more readily, though native speakers may also be better at removing obvious AI wording.

One theory said ChatGPT loves the word "delve" because many of the people rating its answers spoke Nigerian English. Juzek and Ward checked and found no evidence that "delve" and similar words are especially common in any particular variety of English. Where the habit comes from is still open.

So the second idea: a cultulizer. A skill that keeps your English yours. It would preserve the phrasing and references of the writer's own culture instead of converting everything into the same neutral American voice. The 29% result suggests that asking a model to preserve these markers already helps.

Should these tools exist?

A humanizer, a mistakaizer, and a cultulizer can make AI-assisted writing more honest to its author. They can also make AI writing better at hiding.

A 2025 stylometry study from University College Cork found that AI-generated stories cluster tightly together, while human writing is "more varied and idiosyncratic." Maybe the answer is to leave the person in: your opinion, your accent, your occasional typo.

Disclosure does not settle it either. A small 2026 experiment with 34 news readers found that detailed AI-use disclosures reduced trust, while a one-line disclosure performed about the same as none.

I have not decided whether to build the mistakaizer and the cultulizer. Should they exist? Would you use them, or would they cross a line? Tell us on X.

One SKILL.md that makes blogging with AI dramatically easier

Givi Pataridze — Sat, 03 Oct 2026 19:26:28 +0000

You have a topic. You open your AI assistant and start typing.

First you explain your blog: who reads it, how you write, what you covered last month. Then the topic. The draft comes back sounding like every other draft. You ask for a title. Then a summary. Then a search description and some tags. Somewhere in the middle sits a statistic you never gave it and cannot find anywhere. You delete it. Then you open another chat to write the LinkedIn post.

Next week, you start from an empty context again.

goodsender-blog is an open-source skill: one SKILL.md file with its reference notes that replaces that routine with a workflow your assistant follows. Install it, and your next post starts from a workflow instead of a blank chat.

It works in Claude, ChatGPT, and Gemini, and in coding agents such as Claude Code, Codex, Gemini CLI, GitHub Copilot, and Cursor. The skill is published under the Apache-2.0 license, and a GoodSender account is free.

What changes in your routine

	Empty chat	With goodsender-blog
Context	You re-explain your blog every time	It reads your blog first: byline, tags, voice, and topics you already covered
Topic	Whatever you typed	It states the goal and, if you have not picked an angle, offers two or three with trade-offs and recommends one
Facts	Gaps filled from the model's memory	Every factual claim tied to a source or to your word, then checked in a separate pass
Post fields	Title, summary, description, and tags requested one by one	Prepared together and saved in one draft, with a preview link
Social posts	Written later, in another chat	Prepared for each destination you select, before anything is published
Publishing	Copied by hand into each place	Nothing goes out until you approve the actual destinations

The skill keeps its editorial questions to three moments: one message with what only you know, at most one question about claims that still need a source, and one review package at the end. Connecting your account, choosing destinations, and your assistant's own tool confirmations can add a step. Publishing waits for your separate approval.

A workflow, not a prompt

A prompt tells the assistant what you want. The skill tells it how to get there. It carries an editorial process refined over years: how to choose a topic and an angle, how to find and check facts, and how to build the structure around them.

The process runs in twelve stages, grouped here into four phases. Each stage has a gate the assistant must pass before it moves on.

Understand. With GoodSender connected, the assistant reads your blog settings and your most recent posts to learn the voice, byline, and tags. Then it settles the goal: who the post is for and what they should do afterward. It also reads the post type from your request, whether that is an explainer, a tutorial, an opinion piece, a case study, an interview, or a story, and applies the checks that type needs. It asks you once for what it cannot infer, such as what happened, real numbers, dates, and whether you want images.
Write. Evidence is collected in a fixed order of trust. Every major claim in the outline is tagged with its source, or marked as needing one. The draft is written in your blog's voice and toward the goal, then fact-checked in a separate pass.
Package. Title options, summary, description, tags, and any images you asked for are prepared, the draft is saved to GoodSender, and a preview link is created.
Ship. When you ask for it, with the draft or when you plan to publish, copy for each destination you select is prepared and saved without sending. You review everything and approve where it goes.

The skill works for posts on any topic, including fiction and other creative writing, where it checks continuity and any real-world claims instead of sourcing every sentence. It does not need a repository, a particular model, or a paid image generator. Short requests have their own paths: a quick sample skips the interview, and a narrow edit changes only what you asked for.

Facts checked before your name goes on them

When an empty chat is missing a fact, it can fill the gap itself: a plausible number, a study that sounds right, a quote nobody said. Under your byline, that becomes your mistake.

The skill's first rule is that the model's own knowledge is never a source. A factual claim the assistant introduces goes into the post only when something stands behind it, in this order of trust:

You: what you say in the conversation and the files you share
Git repositories you work in or name
Your blog's published posts
Web sources with a URL, cited in the post

A claim without a source is marked in the draft, never quietly asserted. Then a separate fact-check pass, run in a fresh-context subagent where your assistant supports one, reads the draft as if someone else wrote it. It sweeps for words like "most" and "typically" and for every first-person sentence. It returns a ledger: what was verified, what rests on your word, what was hedged, and what was cut.

If anything still lacks a source, you get one numbered message: confirm, source, or cut. Anything you leave unanswered is hedged or removed, so the draft never stalls.

The check works in both directions. If the model's knowledge is out of date or it simply invents something, the claim does not pass. If a reliable source contradicts your own material, the skill tells you once, with the link, and you decide.

You still own the post. What you state or confirm goes in the way you want it. The ledger shows you which parts nobody checked.

All the post fields in one save

A GoodSender post carries more than its body, and each field feeds a different place:

Field	Where it appears
Title	Page heading and search result title
Summary	Lede under the title, blog index card, RSS
Description	Search snippet and social cards
Tags and categories	Public index pages on your blog
Byline	Author line and structured data
Cover image	Post page, cards, and social previews

Through the GoodSender MCP server, the skill saves them together with the draft. The cover is included when you have an image or your assistant can generate one. It proposes the title with alternatives instead of choosing silently. It writes the summary and the description as two different sentences for two different readers. It reuses the tags your blog already has, because a near-duplicate tag splits one topic across two index pages. Images are uploaded through GoodSender and converted to WebP.

The post is saved as a draft with a signed preview link. There are no extra rounds asking for the description you forgot.

Announcements written for each platform

The skill writes separate announcement copy for each destination you select, framed for the people who follow you there. A professional audience may need the practical implication and a concrete example. A short post may need only the problem and the payoff. Each version is checked for new claims and against the platform's length budget, including the link GoodSender adds, and the whole package is reviewed together before you see it.

Destination	What the skill prepares
X, LinkedIn, Bluesky, Mastodon	Announcement copy for that platform, with a link to your post
dev.to, Hashnode	The full article, with a canonical URL pointing back to your blog
Newsletter	A dedicated subject line and summary, followed by a "Read the full post" button

Distribution is saved as a plan, not sent. A connected account is an option, not consent: if you asked for announcements but have not said where, the skill asks which connected destinations to use, and it makes the newsletter a separate, explicit choice. Before publishing, you see the exact destinations and accounts.

When you publish, GoodSender puts the page live and delivers to every destination you selected: announcements, republished copies, and the newsletter email. Sent email cannot be recalled and external posts are not removed for you, which is why the approval comes first. The skill then reports what was published, delivered, pending, blocked, or failed, each separately.

Get started

Download the skill: goodsender-blog-latest.zip. The link always points to the latest release.
Install it.
- Claude: Customize → Skills → Create skill → Upload a skill, then upload the ZIP as downloaded. Skills must be enabled for your account. To update later, download the link again and re-upload.
- ChatGPT: Skills → Create → Upload from your computer. Availability depends on your ChatGPT plan, and uploaded skills do not sync between the desktop app and web or mobile, so add it on each one you use.
- Gemini (web or Mac app): Settings → Skills → Upload. If the ZIP is not accepted, unpack it and upload the folder.
- Coding agents: unpack the ZIP into the agent's skills folder:

mkdir -p ~/.claude/skills && unzip goodsender-blog-latest.zip -d ~/.claude/skills/   # Claude Code
mkdir -p ~/.agents/skills && unzip goodsender-blog-latest.zip -d ~/.agents/skills/   # Codex, Gemini CLI, Copilot, Cursor

Ask for your first post:

"Use goodsender-blog to write a post about [your topic] and leave it as a draft for review."

That is all you need to start. Writing the post does not need anything else. When the draft is ready to save, the skill asks for a GoodSender account and its connection to your assistant, and helps with whichever step is missing. Once a session, it also checks for a newer release and mentions it in passing; it never installs anything on its own.

Bring the topic you have been putting off. This time the routine starts from your blog, not from an empty chat.

Download goodsender-blog · Start your blog

A blog for people who don’t have time to blog

Givi Pataridze — Thu, 01 Oct 2026 19:53:43 +0000

Bring the idea. Make it a post. Keep it going.

You know your subject. You explain it to colleagues, solve problems, make decisions, and learn things someone else would find useful.

Then someone says, “You should write about that.”

They’re right. But writing the article is only part of the job. There’s formatting, images, publishing, the email, and the announcements. Your idea stays in your notes.

GoodSender is now a blog platform built around helping you take that idea further. Work with your AI assistant to develop it, review the result, and publish to your own blog. Prepare the email and social announcements in the same publishing workflow.

Your experience gives the article its value. You get help turning it into something people can read.

Start with something you already know

Think of a question you keep answering. A decision you can explain. Something you tried that changed how you work.

That is enough to start. Open Claude or Codex with GoodSender connected through MCP, the Model Context Protocol, and give it your rough notes:

“Here’s something people keep asking me about. Turn my explanation into a useful blog post. Keep my perspective, ask me where something needs explaining, and prepare a GoodSender draft for review.”

The assistant helps shape the argument and prepares a draft. You add the example it missed, correct the part it misunderstood, and decide whether the article says what you mean. Then you preview it on your blog.

You can stay with the idea while the assistant handles the steps between rough notes and a draft in GoodSender. You can also open that same draft in the editor and work on it yourself.

This article followed that approach. It was drafted with an LLM from our product direction and repository, revised through conversation, and prepared for publication through MCP.

Start your blog. Bring one idea you have been meaning to write about.

Prepare the post and the invitation to read it

Once the article is ready, there is another piece of writing to do: give people a reason to open it.

GoodSender connects the original post to the different forms that invitation can take:

Email: a summary that gives subscribers a reason to read, followed by a “Read the full post” button.
LinkedIn and X: announcements tailored to each platform, with a link to the article.
Bluesky and Mastodon: announcements for the readers who follow you there.
dev.to and Hashnode: the full article, with a canonical URL pointing back to the original on your blog.

That last detail matters when the same piece appears in several places. The canonical URL identifies your original as the preferred source.

Connect the accounts you want to use, select destinations for the post, and review the distribution copy before publishing. A personal essay and a practical tutorial can reach their readers through different channels.

Email is built into GoodSender, backed by our sending infrastructure on Laneful. Subscription confirmation, unsubscribes, and complaint handling are part of that foundation.

Give the next post a starting point

An article often leaves you with another question worth exploring. You explained an idea. Next, show how you used it. Then write about what you learned.

Ask your assistant to propose a series from the material you already have. A set of research notes, lessons from managing a team, or stories behind product improvements can each give you several connected articles. A changelog is another possible source.

You can also give your assistant access to the repository you work on as a developer, tester, or owner, or several repositories you work across. Ask it to find the stories in the code, documentation, and changes, then help you explain why they matter.

With a scheduled task in Claude Cowork and GoodSender MCP connected, you can make preparing the next draft a recurring part of your routine. For example:

“Every Friday, review the new notes in this folder. Suggest the next article in our series and prepare a GoodSender draft. Ask me about missing examples and leave it ready for my review.”

The recurring task runs in Cowork. GoodSender holds the drafts and publishes the posts. You review each piece, decide what is ready, and can schedule finished posts for publication.

The next writing session begins with something to respond to.

Ready for readers, search, and AI

Your idea deserves a page that is ready for readers.

Your blog starts with templates already optimized for Performance, Accessibility, Best Practices, and SEO. That foundation is ready to use: see the scores from a test of a GoodSender blog below.

Underneath the page, GoodSender handles rendered HTML, metadata, canonical tags, a sitemap, and RSS. Search crawlers and AI systems can read the article text directly. Those are the foundations behind its SEO and answer engine optimization, or AEO.

Images get attention too. Uploaded images are converted to WebP and prepared in multiple sizes, taking image conversion and resizing out of your publishing checklist.

Your blog lives on your own domain. Email and social announcements bring readers back to the same growing collection of your work.

Coming next

We are working toward fully customizable blog design through AI conversation, while preserving the SEO and AEO foundations built into our templates. Describe the blog you want, then refine its design by talking through changes. This feature is upcoming.

For your first post, start with the thing you explained to someone this week.

Create your GoodSender account, connect your domain, and turn that explanation into an article. You can begin drafting while your domain is being connected.

Start your blog

A $169 DIY AC controller that saves 450 kWh a month

Givi Pataridze — Thu, 24 Sep 2026 20:59:44 +0000

A friend of mine runs a small photo studio, rented out by the hour. There is no staff. Photographers book a slot, punch a door code, and let themselves in with their clients and their gear.

That last part is the whole problem. You cannot ask a paying client to stand around in a hot room while a wall-mounted air conditioner catches up. And a room full of people under shoot conditions goes stale long before it gets hot, which is why this story involves a CO2 sensor.

Before this project, the climate was run by whoever was there last. They set the air conditioners with a remote and left. So the units routinely ran all night for an empty room, and just as often the next booking walked into the wrong temperature anyway.

My friend is not there most days. I am almost never there. But I am the person in our friend group with a soldering-adjacent reputation, so I got volunteered.

I had never actually soldered anything. That becomes relevant.

the plan that was supposed to cost $60

The studio runs on a small budget. That is why its two air conditioners are cheap white-label Chinese splits rather than a known brand, and it is why the proper routes, a commercial climate-automation install or one of the smart-AC-controller gadgets, were never seriously on the table. They cost several times more than what I ended up spending, and what I ended up spending was already too much.

On July 17 I asked Claude to pick the parts. The requirements: a small ESP32 controller board, Ethernet with power over the cable, a sensor for CO2, temperature, and humidity, and infrared control for the two ACs. Infrared control means teaching a circuit board to click the remote electronically. The AC cannot tell the difference between a remote in a human hand and a small LED blinking the same pattern at it.

The first estimate came back at about $60.

the cart that checked out

By July 20 the plan had grown. Instead of small boards hidden near each AC, I wanted one elegant wall-mounted device with a color screen: temperature in big digits, CO2 next to it, buttons on the front. The cart, all from M5Stack's shop:

Part	Price	Role
M5Stack Basic v2.7	$39.90	`ESP32` controller with the color screen
LAN PoE base W5500 v1.2	$48.50	Ethernet plus power over the network cable
Unit CO2L (`SCD41`)	$45.90	CO2, temperature, humidity
2x Unit IR	$7.90 each	infrared emitters, one per AC
Grove cables and jumpers	$10.40	wiring

Total: $169.70 (including shipping). Claude confirmed the cart item by item. Its verdict: checks out, safe to order.

Every item on that list was correct. The cart was fine.

It just wasn't the plan.

what arrived, and what didn't fit

The boxes arrived August 13, and every part was individually fine. The catch showed up during assembly, and it was not a missing part. It was a budget problem of a different kind: one small ESP32 does not have enough free pins and communication channels to run everything at once.

Think of the board's pins as outlets on a power strip, except some devices refuse to share and some insist on a specific outlet. The Ethernet chip in the PoE base claims the SPI bus plus a handful of GPIO pins. The color screen needs that same SPI bus plus pins of its own. The CO2 sensor takes the I2C port. And each infrared unit wants its own direct GPIO line, because infrared needs exact microsecond timing that can't be routed through a shared connection.

Ethernet, screen, sensor, and two IR channels together simply do not fit on this board.

Something had to lose.

what got cut

Ethernet lost first. The $48 PoE base stayed in the build, but only as a power supply: the network cable delivers electricity and nothing else, and data moved to Wi-Fi permanently. There is a comment in the shipped config that says exactly that. I paid $48 for a power brick.

The IR receive channels lost second. The infrared units run transmit-only; I wired up a receiver when I needed to record the real remote's signals.

Then came the second catch. The screw terminals on the PoE base, the natural exit point for the IR wiring, are not connected to anything from the factory. Power, ground, and the control pins simply do not reach them. The usable pads exist only inside the stack. Claude and I walked through every no-solder alternative, and each one died. A hub is the wrong kind of connection: IR needs a direct GPIO with a 38 kHz carrier and microsecond timing, and every hub in this ecosystem is an I2C device that physically cannot do that. A $6 pass-through module needs soldering to attach anyway. Claude's summary: the six terminals are a blank to be soldered, not a ready breakout.

I had specified this build as no-soldering. Every path led back to a soldering iron. This is the point where I swore at my AI assistant, and its answer boiled down to: the parts are all fine, the sentence missing at order time was that wiring the IR units out of this stack means fifteen minutes with an iron.

It was right, which did not help.

four wires

So I bought a soldering iron, the first one I have ever owned, and practiced on a dead board.

The actual job was four jumper wires from the base's internal pads to the screw terminals: ground, 5V, and the two IR signal lines, G16 and G17.

It worked.

Four wires. That was the entire distance between one elegant device with a screen and a stack of parts that talk past each other.

an air conditioner that spoke no known language

Now the software half. The ACs are branded "Wagner," a regional white-label name, remote model G4W-18BHR. The obvious theory is that a no-name split is one of the usual suspects inside: Gree, Midea, the Coolix protocol family. ESPHome's IR climate support covers all of them out of the box. You capture one button press, the library names the protocol, and you're done.

I got an early false victory. The first remote I captured, to test the receiver wiring, was my home AC's remote. It decoded instantly as Coolix. For a while the plan was: it's Coolix, config done.

The actual remote was not the same remote.

The Wagner remote matched none of the roughly ten ready-made protocol decoders I tried. Gree in all its variants, Coolix, Midea, TCL112, everything. Nothing recognized it. Somewhere a factory had shipped an air conditioner speaking an infrared dialect none of the usual open-source decoders recognized.

reverse-engineering the ir protocol from 97 recordings

So we decoded it from scratch: me pressing buttons, Claude staring at the timings. I captured 97 infrared frames, and on September 1 the pattern gave in.

The result, now documented in the project repo: a 38 kHz carrier, which is the specific flicker rate the AC's receiver eye is tuned to. A 96-bit frame, 12 bytes, sent least-significant-bit first. A 6180 microsecond header pulse announcing each message. Temperature encoded as the value minus 16 in byte 9, so 24°C travels as an 8. Fan speeds numbered backward, 1 for max and 3 for min, because of course they are. And a checksum, the frame's built-in way of noticing corruption, that is nothing fancier than the sum of the first 11 bytes modulo 256.

No existing ESPHome component spoke it, so we wrote one: a custom decoder and encoder, built largely with Claude doing the byte arithmetic and me doing the button pressing.

proving it

A decoded protocol is a theory until it round-trips. We re-encoded every captured frame from the formula and compared the output byte for byte against the recordings. The first run scored 44 of 45; the single mismatch was a transitional frame caught mid-mode-switch. The control run scored 52 of 52.

Then the loop closed itself in the most satisfying way available. On the first live transmission, the device's own receiver caught the frame its own transmitter had just sent, and decoded it correctly. The bytes matched a recording of the real remote for the same state exactly.

On September 4 the real indoor unit obeyed: on, off, factory turbo.

The protocol was closed.

the bug that wasn't

One debugging story worth its space. The IR LED on the emitter module is fed through a 150 ohm resistor, about 23 mA, against the hundreds of milliamps a real remote pushes. Range is roughly two meters and the beam is narrow.

One day the AC stopped responding. I burned hours on a plausible hypothesis: maybe a single frame wasn't enough and the command needed repeating, the way some remotes hammer their message several times. Four-copy frames worked once, by coincidence, then failed too. The actual cause, the entire time, was that the LED was pointed slightly off axis. Aimed properly, a single frame was enough, exactly like the factory remote.

With a weak emitter, aim matters more than distance. The AC was never ignoring me. I was talking past it.

controlling a split system remotely, without the remote

Here is the finished behavior, running on the wall with nobody in the building, no server and no phone app required to keep it running.

The room sensor is the thermostat, not the AC's own one, and it updates every 60 seconds. In summer, cooling switches on at 24.0°C and off at 23.0°C. In winter, heating switches on at 18.0°C and off at 22.0°C.

The trick that makes a cheap split behave is the setpoint. The box deliberately commands the AC to an overshoot target, 18°C when cooling and 24°C when heating, so the AC's own thermostat never decides it's done and starts dawdling. The box itself cuts the unit off when the room sensor says done. The AC works at full tilt or not at all, and the room, not the machine, is the judge.

Every 15 minutes the box re-sends its current command. Anyone who fixes the climate with a physical remote gets quietly overruled within a quarter of an hour. The remotes still work; they just don't win.

The night schedule is the money part. Both units switch off from 23:00 to 08:30, every day, because night bookings are rare and an empty room does not need conditioning. A frost guard heats if the room falls to 10°C and stops at 13°C. The rare night visitor presses a button on the box to suspend the schedule.

Three buttons total: cycle auto/manual/off, toggle the night schedule, and a 180-minute boost at the coldest setpoint and max fan. The screen shows temperature in half-screen digits and CO2 color-coded: green below 1000 ppm, yellow to 1500, red above. Red means open a window, which in a room full of people under shoot lights happens well before anyone feels warm.

And a failsafe: if the temperature sensor drops out, the box switches both ACs off rather than run blind. Turbo and off are the only commands it will still send without sensor data. There is no server, no cloud, no app. If Wi-Fi dies, the logic keeps running on the device.

turning the ac off at night: what it saves

I did not install an electricity meter, so what follows is arithmetic on stated assumptions, not a measurement. The assumptions are in the open; disagree with any of them and the number moves.

The remote's model number marks these as 18,000 BTU-class splits, which draw roughly 1.5 to 1.8 kW while the compressor runs. Left on overnight, a split's compressor cycles; call it running about half the time. The schedule removes 9.5 hours a night, on two units.

Two units, 9.5 hours, about 1.65 kW, half duty: roughly 15 kWh per night, with a fair range of 12 to 19. Per month, roughly 450 kWh, range about 350 to 570.

Pricing is illustrative arithmetic only: at $0.10 per kWh that is about $45 a month, so the $166.40 of hardware pays for itself in around four months on the night schedule alone, before counting whatever the daytime hysteresis saves. Your tariff will differ. The shape of the number will not: the biggest feature in this build, the one that actually saves electricity, is the off switch.

the same build, cheaper: no screen, no PoE

Now the honest postscript. Knowing everything above, here is what I would actually build.

The roughly $50 version, no soldering. Two M5Stack Atom Lite boards, under $10 each, one stuck next to each AC. The IR LED is already on the board, so the entire soldering saga in this post evaporates. Add one CO2 sensor unit on the Grove port of one Atom (the SCD40 version was $33.90 when I checked in July; the SCD41 in my build costs more). Power both from any USB phone charger. Run the same ESPHome logic over Wi-Fi. No screen: the numbers live in ESPHome's built-in web page, readable on any phone. No PoE: the phone charger is the power supply. This is the version I should have started with.

The roughly $20 version, if you can solder a little. Any generic ESP32 dev board, $4 to $6. An IR LED plus a transistor and two resistors, under a dollar in parts, is the whole IR side. A basic temperature and humidity sensor, a few dollars, covers the thermostat. That gets you the hysteresis, the overshoot trick, and the night schedule: everything in this post except the CO2 number.

Where not to cheap out: the CO2 sensor. A real CO2 sensor, photoacoustic or NDIR like the SCD40 and SCD41, is the one part that costs real money, $30 and up, and it costs that for a reason. Cheaper boards advertise eCO2, which is an estimate derived from other gases, not a CO2 measurement. That warning surfaced in my July parts research and is why the SCD41 is in the box. If the room-full-of-people number matters to you, pay for it. If it doesn't, drop it, and the whole build gets very cheap.

What did the $169.70 actually buy over the $50 version? A screen on the wall, so the numbers are visible without a phone. Three physical buttons. And $48 of Ethernet that carries no data. The logic, the protocol work, and the savings are identical on the cheap builds.

I would build the $50 version today. The $169.70 version is what it cost me to learn that.

Because here is the thing the whole project turns on: not one expensive part ever failed. The board was fine, the base was fine, the sensor was fine, and Claude's cart really did check out, line by line. What nobody counted was the connections: the shared bus, the pin budget, the six terminals that were a blank.

BMAD Ate My Month: When Spec-Driven Development Is Too Much

Givi Pataridze — Tue, 22 Sep 2026 08:20:39 +0000

On September 5 I deleted 1,782 files from my autopilot's repository in one commit. Agents, templates, workflow configs, a PRD (Product Requirements Document), an architecture document, a story backlog longer than most novels. About 16 MB of process, and not one line of product.

Typing that commit message felt better than anything I shipped that month.

Some context if you are new here. The autopilot is an autonomous coding system I built and have been running for months: it receives GitHub webhooks and runs headless Claude Code sessions that write, review, and merge pull requests without me at the keyboard. I have told its story before, from building it to the day it ran out of tokens. This post is about the question those posts skipped: how work should be planned before an agent ever sees it.

The industry has an answer to that question. SDD (Spec-driven development): write the specification first, make it the contract, and let the agents implement against it. BMAD is that idea at full weight, an entire agile organization in software, with AI analysts, architects, product managers, and developers, and a workflow that runs them in order. For four weeks, it planned my product.

This post is about what that month cost, and the embarrassingly small thing that actually fixed the problem BMAD was supposed to fix. If you run coding agents and your PRs keep turning into hostage negotiations, this is for you. I got this wrong twice before getting it right, and both wrong answers looked smart at the time.

The failure that started everything

You know this failure. You have probably committed it this week.

You are busy, the agent is capable, so you file a one-line issue: remove this, add that, make it configurable. The agent reads your line and builds either the wrong thing or a monument to the right one. Then the reviewer model, the expensive one you hired precisely to catch problems, catches the problem. In review. After the work exists. After the money is spent.

My worst case: a cleanup that should have been a quiet deletion collected sixty comments across ten review rounds. Two models arguing with each other, round after round, every round billed to me. Then I closed the pull request without merging it.

Sixty comments. Ten rounds. Nothing shipped. And the whole argument was downstream of one line I wrote in thirty seconds.

Attempt 1: the machine tried to fix itself

My autopilot's answer to "the PRs are too big" was, I swear, more autopilot.

It generated a size gate: a scope agent that estimated the diff an issue would produce, and above a budget carved the work into ordered sub-issues. Almost six thousand lines of new code, to enforce a line budget on future code. I did not design this. The machine did, and the design was even accurate about the disease. Its own PR body diagnosed that large PRs drive the review churn.

And still it was attacking the symptom. The gate measured work after the intent was already written badly. Splitting work after it exists is triage, not planning. By the time there is a diff to estimate, the mistake is old. It was made in the issue.

I never merged it. The dangerous part is how reasonable it looked.

Attempt 2: BMAD

I found BMAD through a podcast on spec-driven development. One of the guests was a BMAD contributor, and everything he said sounded solid. I wanted solid. I was tired of arguing with my own pull requests.

So the day after rejecting the machine's six-thousand-line fix, I installed a 1,782-file one.

BMAD planned my product the way an enterprise plans a satellite launch. A 13,397-word PRD. A 19,867-word architecture spine. An epics.md backlog that reached 43,062 words. Nine epics, 86 stories, every one with acceptance criteria. Reading the plan felt like adult supervision had finally arrived.

Five days in, an implementation-readiness assessment found one critical omission and three major planning defects, and blocked all further implementation. Sit with that. The plan failed a review of the plan. The fix, approved the same day, grew the plan to 91 stories. The product had not moved.

Then the loop started, and I met the real bill. A typical story burned one to two million tokens. One story hit around thirty million. Thirty million tokens, one story. Maybe I misconfigured something; honestly, I probably did. But look at what my config had to become in self-defense: a four-million-token budget per story, a 210-minute session timeout, a hard cap of seven review cycles, nudge budgets for sessions that stall without a result. Those are not settings. Those are restraining orders.

To be fair, bmad-loop itself is genuinely well built. The TUI is impressive. The quality bar is real; the reviews caught real defects. In four weeks, 23 stories were implemented and reviewed through it. Six more finished their code and parked themselves as "awaiting operator", waiting for things only a human can do.

But my product was young and changed constantly, and every change created homework. Update the PRD. Update the architecture. Update the backlog. The documents needed to be current before the code was allowed to move, so the code moved at the speed of the documents.

I was supposed to operate an autonomous system. I had become its secretary.

My verdict is not that BMAD is bad. BMAD is solid. Too solid. For an established product with a stable spec, that bar might be exactly right. For a product still finding its shape, the bar killed the only metric that mattered: time to release. On September 5 the whole kit left in one commit, and the release line moved forward the same day.

Attempt 3: steal the good part, skip the religion

Here is what actually worked, and what I would tell you to do instead of either of my failures.

I did not adopt another framework. I took the one thing from the spec-driven world that had earned its keep, Spec Kit's way of expressing intent, and forced it into add-autopilot-issue, the skill I already used to file work. Spec Kit is compact where BMAD is encyclopedic, and it asks quite exactly the right questions: what changes, for whom, how you will know it worked.

The mechanics that now run my intake, all of them small:

Every story issue carries a pinned spec reference: Spec: owner/repo@sha:path#user-story-N, full 40-character SHA required. Coder and reviewer read that exact version, so neither can move the goalposts mid-session by editing the spec.
One user story becomes one issue. Two or more become a parent epic issue with one issue per story and an epic/NNN-<slug> integration branch.
A story with more than 12 tasks gets split before it is filed.
Reviews run against the pinned spec with a BLOCKING/SHOULD/NIT rubric, so "I don't like it" and "this violates the spec" stop being the same comment.

And one rule no code enforces, because it is mine: when I write an issue, I aim for work around 300 lines of non-test code. Small enough that a session cannot wander, small enough that a review cannot spiral.

That is the entire fix. A reworked skill and a pinned line of text. The smallest thing I tried, and the only one still on main.

Decided before it starts

Three attempts, one lesson wearing three costumes. The quality and cost of an autonomous coding session are decided before it starts: by how precisely the intent is expressed, and by how small the work arrives. The size gate tried to enforce that in code, after intent. BMAD tried to enforce it in documents, on top of intent. What worked was just doing it: saying precisely what I want, in pieces small enough to survive contact with review.

One more observation, offered as observed because I have not measured it and it is clearly not linear: the better the intent is expressed, the cheaper the model that can execute it.

And a verdict, for whoever is weighing BMAD right now. I did not dabble. I ran the full, fundamental approach, the way it is meant to be run, and for me it did not work: too long, too expensive. For an enterprise with a stable spec and a tolerance for ceremony, that same weight may be exactly right. My projects move fast, and BMAD could not keep up with them. So the experiment is closed. I will not run it on any of my projects for now; more flexible approaches carry its best idea at a fraction of the cost.

That is what those 1,782 deleted files were. Not a failed tool.

A closed experiment.

How to Send Email via Claude Desktop

Givi Pataridze — Wed, 10 Jun 2026 08:23:17 +0000

How to Send Email from Claude — Free, in Two Minutes

Two minutes from now, Claude Desktop will send a real email through your own account. This guide is for Claude Desktop users with a GoodSender account who want to wire up email via MCP. GoodSender is free for the first 100,000 emails per month, no credit card required.

Prerequisites

Claude Desktop installed.
A GoodSender account — sign up free, no credit card.

Step 1: Install the MCP Server

Download goodsender.mcpb from the latest GitHub release. Double-click it. Claude Desktop registers the server automatically.

Using a different AI client like Cursor? You'll need the Docker or manual binary path — see the MCP server docs.

Step 2: Add Your API Key

Grab your API key from the GoodSender dashboard. Paste it into the configuration prompt Claude Desktop shows after install.

Before sending, make sure you've also verified your sending domain in the GoodSender dashboard.

Flow 1: Compose and Send a Custom Template

In Claude Desktop, ask Claude to create a "weekly update" template using the MCP template composer:

Create a new email template called "Weekly Update" with a subject line "Your weekly update from Awesome App" with a futuristic design using GoodSender MCP

Claude opens the MCP template composer and builds the template with your variables. Once you're happy with it, ask Claude to send it:

Send the Weekly Update template to me at you@example.com.

A note on consent: Custom sends to recipients require the Permission Loop. GoodSender requests consent from the recipient first — outbound messages queue locally until consent is granted. Once a recipient approves, subsequent sends from your domain go through instantly. That earned permission is what keeps deliverability compounding rather than degrading over time.

Flow 2: Send a Transactional OTP Email

Transactional templates are the contrast case — no template setup, no consent step, and they fire instantly. In Claude Desktop, paste this prompt:

Send an OTP code to me at you@example.com.

Claude picks the GoodSender MCP, calls the transactional template endpoint, and the email lands in your inbox instantly. Transactional templates ship without a consent step, so there's no queue and no setup — the send goes through the moment Claude calls the tool.

Flow 3: Schedule a One-Time Send

Claude Desktop (in Cowork mode) has a built-in scheduler. You can use it to schedule a one-time send of the Weekly Update template for Monday at 10am — GoodSender's MCP handles the actual send when the scheduled time fires.

Schedule a one-time send of the Weekly Update template to you@example.com for Monday at 10am. Use headline "Monday briefing" and "Here's what's coming up this week."

What You Can Do Next

Now that you've got email working from Claude, the best thing to do is try it with your own use case. Send an OTP to a real user, schedule a weekly update, or write an email in Markdown and watch it render perfectly across 90+ inbox clients. The setup is already done — the rest is just asking Claude.

Why GoodSender for This

Free for 100,000 emails per month, no credit card. Consent and engagement are built into the API — the Permission Loop gates custom sends on explicit recipient consent, and the Engagement Check monitors recipient activity over time — so deliverability compounds rather than degrades. It runs on the same IP infrastructure Laneful uses for the world's largest senders.

Get Started

Get your free API key →

How to Improve Email Deliverability in 2026: Stop Fighting for the Inbox and Start Earning It

Givi Pataridze — Tue, 02 Jun 2026 19:39:02 +0000

You did everything right. You set up DKIM. You verified your domain. You wrote a clean SPF record, layered DMARC on top, and kept your recipient list tidy by hand. And your password-reset emails still land in spam.

The instinct, at this point, is to look for another DNS record to fix. To re-read the Postmark blog. To check whether the alignment mode in your DMARC policy is relaxed when it should be strict. That instinct is wrong — or, more precisely, it ran out of useful work several steps ago.

Deliverability is not a configuration problem. It is a reputation problem. And reputation is not something you configure once and forget; it is something you earn, continuously, through the behavior of the people who receive your mail. The authentication checklist is the floor. Everything that actually decides whether you reach the inbox happens after the SMTP handshake.

Why the Standard Advice Isn't Enough Anymore

The checklist is necessary. It is also not differentiating. DKIM, SPF, DMARC, a custom sending domain, sane bounce handling — every credible email provider can set this up for you, and many will do it automatically. That is not the interesting part. Mailbox providers treat authentication as a precondition for being considered at all, not as a signal of trustworthiness. A spammer can pass DKIM.

The more uncomfortable problem is structural. Most email API customers share IP pools with thousands of other senders, and they do not get to choose who those other senders are. Some of them have poor list hygiene. Some buy lists. Some are running outright abuse campaigns until the provider catches them. While they are sending, every sender in the pool pays — through higher spam placement rates, throttling, and occasional blocks. Everyone's reputation rises and falls together, so good senders end up paying for bad senders' mistakes.

This is not an accident of how the market evolved. It is baked into the pricing. SendGrid's Essentials plan starts at $19.95/month (up to 100,000 emails); serious 100K usage typically pushes customers onto Pro at $89.95/month. Mailgun charges around $75 at 100K. Postmark sits at roughly $134. Resend lands at $35. Those prices, in part, absorb the cost of cold-email abuse across the shared pool — complaint handling, reputation remediation, abuse-team overhead. Every customer pays for those problems whether or not they cause them.

The 2026 context makes this worse, not better. Gmail and Yahoo tightened bulk sender requirements in February 2024 — for senders sending 5,000 or more messages per day to personal Gmail or Yahoo accounts, one-click unsubscribe headers (RFC 8058), stricter authentication enforcement, and a spam-complaint rate kept under roughly 0.3% are now mandatory. Google escalated enforcement through 2025, moving from temporary delays toward outright rejections for repeat offenders. The direction of travel is toward stricter enforcement, not looser. Passing authentication is increasingly the minimum bar to avoid outright rejection — not a path to the inbox.

What Mailbox Providers Actually Reward

The positive case is worth stating directly, because it changes what you optimize for.

Mailbox providers care primarily about engagement signals. Opens, clicks, replies, and "move to inbox" actions tell them that recipients want the mail. Spam complaints, deletions without opening, and "mark as spam" actions tell them the opposite. Reputation is a running score, not a one-time certification — it compounds in both directions, and the input is the behavior of your recipients.

Complaint rates are the fastest way to damage that score. A single spam complaint from someone who never affirmatively asked to hear from you is disproportionately damaging, and you cannot rely on feedback loops alone to catch the problem. Gmail does not run a traditional FBL; the protection mechanism there is aggregate reputation surfaced through Google Postmaster Tools. By the time a complaint registers as a number on a dashboard, the damage is already underway.

List decay is the slower, quieter killer. Recipients who stop engaging but stay on your list become a liability. They don't complain. They just don't open. Sending consistent volume to recipients who never engage is itself a signal that you are not sending wanted mail. Most senders never notice this happening, because their email API surfaces engagement data in a dashboard nobody opens — not as something the system acts on.

The Three Levers That Actually Move Deliverability

What follows is a principles-first playbook. Each lever can be implemented manually on any platform; the question is whether you want to build it yourself or use infrastructure that enforces it for you.

Lever 1: Permission before the first send

The highest-leverage thing you can do for deliverability is to make sure every recipient on a custom send list has explicitly consented to receive mail from you.

This is not the same as "they gave us their email address at signup." It means the recipient has affirmatively opted in to receive this type of communication from this sender, and you have a record of it. Mechanically, a consented recipient is far more likely to open, click, and not complain. Every send to a consented list is a positive reputation signal. Every send to a list you assembled from form submissions and CRM exports is a roll of the dice.

The hard version of this principle is enforcing consent at the API level — not as a best-practice reminder in the docs, but as a physical gate. The sender calls a consent endpoint, and the system delivers a short, high-reputation consent email to the recipient with approve and reject buttons. The link is signed and cannot be forged. Until the recipient clicks approve, any custom send attempt to that address is discarded — not bounced, not deferred. Because every recipient is consented, complaint rates stay low by construction.

If you are sending only OTPs and receipts, this section feels inapplicable. It mostly is. Transactional templates — MFA enrollment, OTP codes, order confirmations, new-device alerts — send without the consent step. But the infrastructure-quality argument in Lever 3 still applies to you, and the next section explains why.

Lever 2: Continuous engagement monitoring with automatic suppression

Permission at signup is not enough. Recipients disengage over time, change jobs, abandon inboxes. A list that was healthy six months ago is a reputation liability today if you haven't been watching it.

The fix is continuous monitoring — tracking opens, clicks, bounces, and complaints per recipient, and acting on the data automatically. Three specific behaviors matter:

Re-permission disengaged recipients before they hurt you. An automatic re-permission ping after a period without engagement keeps lists clean. If the recipient doesn't respond, the address drops off the list.
Suppress complaints and unsubscribes instantly, workspace-wide. When a recipient unsubscribes or marks as spam, the API should physically refuse subsequent sends to that address from any key in the workspace — not later, not after the next list import, immediately.
Drop hard bounces on the first hit. Sending to addresses that no longer exist is one of the strongest negative signals available.

Most email APIs surface these metrics in a dashboard. Few enforce them automatically. The difference matters in practice, because a developer who has to manually review engagement data and manually update suppression lists will not do it consistently. Nobody does.

Lever 3: Infrastructure that doesn't punish you for other senders' behavior

Even with perfect permission and engagement practices, a sender on a contaminated shared IP pool faces headwinds that have nothing to do with their own behavior. So the infrastructure question is unavoidable: who else is in your IP pool, and what are they doing?

The alternative to a generic shared pool is a pool where every sender is operating under the same permission-and-engagement constraints — so the aggregate complaint rate stays low by design rather than by luck. When a shared pool is built this way, it is already warm; new senders can send at full volume from day one without a manual warm-up period. At higher volumes, moving to dedicated IP allocation removes the shared-pool variable entirely.

The Transactional Email Caveat

If you are reading this and thinking "I only send password resets and OTPs, this doesn't apply to me," part of that is correct and part of it is not.

Transactional senders often assume they are safe from deliverability problems because their mail is, by definition, wanted. The recipient just asked for a password reset; of course they want the email. That logic is partially true. It is also incomplete.

Transactional email still lands in spam if the sending IP has poor reputation from other traffic in the same pool. The recipient's intent doesn't matter to the spam filter — the filter never sees it. What the filter sees is the IP, the domain, the authentication result, and the historical pattern. If those signals are bad, your password reset goes to spam even though every recipient genuinely wants it.

Bounce handling matters here too. Sending OTPs to addresses that no longer exist degrades reputation for the addresses that do. And the moment you add any custom or marketing sends to the same workspace, the reputation of that traffic flows back into your transactional deliverability. Most products start transactional-only and grow into custom sends eventually. Choosing infrastructure that enforces permission on the custom side from day one prevents that growth from quietly degrading the transactional path.

Why Most Email APIs Make This Hard by Design

It is worth asking why permission and engagement tooling is so rare in the developer email market. The answer is structural, not accidental.

The legacy pricing model — SendGrid Essentials at $19.95/month (up to 100K emails), Pro at $89.95/month for serious 100K usage, $35 at Resend, $75 at Mailgun, around $134 at Postmark for 100,000 emails — is calibrated to a world where the shared pool absorbs the cost of abuse. If a provider's revenue model depends on the volume that includes senders with poor practices, there is limited economic incentive to enforce permission at the API. Doing so would shrink the addressable market and reduce the volume that justifies the price.

The result is that none of the major developer email providers ship permission verification as a product feature. Mailgun has none. Resend has none. Postmark has none. Mailtrap has none. Amazon SES, at roughly $10–$12 per month for 100,000 emails, is cheaper, but the trade-off is an AWS account, manual IP warm-up, and no permission management of any kind. A developer who wants to implement permission correctly on any of these platforms has to build it themselves — consent tracking, signed approval links, suppression lists, re-permission flows. It is non-trivial engineering work that has nothing to do with the product they are actually trying to ship.

When permission is enforced at the API level, the economics change. Complaint rates stay low by construction. Sender reputation compounds positively. Abuse-team overhead collapses. The cost structure that justified higher monthly pricing no longer applies.

This is the structural shift worth paying attention to. The market is the way it is because of how shared pools and shared pricing reinforce each other. Break the loop on either side and the rest of the model becomes optional.

The Earned Deliverability Playbook

To compress the argument into something you can act on:

Authenticate your domain. DKIM, SPF, DMARC. Required, not differentiating. Do it once, correctly, and move on.
Gate custom sends on explicit consent. Do not send marketing or custom email to anyone who has not affirmatively approved it. Enforce this at the API level if you can, not as a manual process you mean to get to.
Monitor engagement continuously and suppress automatically. Re-permission disengaged recipients before they become liabilities. Suppress complaints and unsubscribes instantly, across every API key in the workspace.
Choose infrastructure where the pool is protected by design. Your deliverability depends partly on who else is sending from your IP range. That is a choice, not a fixed cost.
Treat transactional email as a deliverability asset, not a given. OTPs and receipts land in spam too if the infrastructure around them is contaminated.

Deliverability is not a problem you solve. It is a reputation you build — through permission, engagement, and infrastructure that compounds both. The largest senders in the world have known this for years; the playbook is not new. What is new is that it can be automated.

Earned deliverability, automated. GoodSender was built to handle exactly this — the permission, engagement, and infrastructure discipline described above, enforced by design so you don't have to build it yourself.

Cold Email's Math Broke in 2026. Here's What Replaces It.

Givi Pataridze — Tue, 26 May 2026 17:49:10 +0000

Cold email has always made a simple promise: buy a list of contacts, set up an automated sequence, and new sales conversations start landing — cheaply and at scale. For a long time the math more or less held. In 2026 it doesn't, and the reason isn't your copy or your sending tool. It's that the inbox has quietly changed the rules, and cold email is now playing a much harder game, with much less room for error.

If you're a founder weighing how to spend the next quarter of outreach budget, this is worth understanding before you commit to another sequence.

The short answer, before the why: the more effective replacement for cold email is permission-based sending — reaching a smaller list of people who opted in and stay engaged, because that's exactly what the 2026 inbox rewards.

Cold outreach and permission-based sending chase the same goal — but the 2026 inbox rewards only one of them.

What actually changed

Starting in 2024, Gmail and Yahoo began enforcing real requirements on bulk senders — for Gmail, anyone sending 5,000 or more messages per day to personal Gmail accounts, a classification that becomes permanent once you cross it. Microsoft joined them in May 2025. The headline rules are authentication (SPF, DKIM, DMARC) and one-click unsubscribe, but the one that quietly changes the math for cold email is the spam-complaint threshold. Google's own guidance is to keep your complaint rate below 0.1% and never let it reach 0.3%; at 0.3% or higher, bulk senders lose access to Gmail's mitigation support and take an even stronger hit to inbox delivery, and non-compliant mail can be rejected or sent to spam outright.

Google asks bulk senders to stay under 0.1% spam complaints and treats 0.3% as the danger line.

To put that in human terms: at just one complaint per 1,000 emails you're already past the 0.1% Google wants you under, and at three per 1,000 you've hit the 0.3% line where Gmail pulls its mitigation support. Cold lists are structurally more likely to approach that line, because the recipients never explicitly asked to hear from you. Some of them will be annoyed, and a small fraction of annoyed people is all it takes.

Here's the part most teams discover too late. The damage isn't always contained to the campaign. Mailbox providers attach reputation to your sending domain and IP and weigh it on signals like complaints, bounces, engagement patterns, authentication, sending consistency, and list quality. A cold campaign that generates complaints doesn't just underperform; it can weaken the reputation signals attached to the domain, subdomain, or infrastructure you send from. If your promotional and transactional mail aren't separated properly, that risk can spill over into messages you actually need people to receive: password resets, receipts, onboarding emails, and real customer replies. The cheap channel turns out to carry one of the most expensive failure modes there is.

The real problem is permission, not tactics

The instinct is to respond with better tactics: warm-up tools, more inboxes to rotate through, sharper subject lines, AI-written personalization. The harder problem is who you're up against. Mailbox providers read signals across billions of messages and keep getting better at spotting mail people didn't ask for, so tactical workarounds tend to have a short shelf life — and the usual cost of leaning on them is burned sending domains.

The more useful move is to notice what the largest, most sophisticated senders in the world actually do. They don't lean on cold email. They build permission and engagement into the system from the start, because that's what the inbox most consistently rewards. Permission gets you through the door; sustained engagement keeps it open. Everything else is a workaround for not having those two things.

That reframes the question. The goal was never "how do I send more cold email without getting caught." It's "how do I build a list of people who actually want to hear from me, as efficiently as cold email promised to be." That list is the asset. It converts better, it complains far less, and it protects the deliverability of your whole domain instead of eroding it.

What the replacement looks like in practice

A permission-first outreach motion has three moving parts, and none of them are exotic.

First, you ask before you send. Instead of dropping a pitch into a stranger's inbox, you send a single lightweight message that lets them opt in — one clean ask, with a clear yes. The people who say yes are now reachable; the people who ignore it never enter your promotional sending stream, so they can't complain about a follow-up sequence they never agreed to receive.

Second, you watch engagement and act on it. People who stop opening and clicking get re-confirmed or quietly dropped, so your active list stays genuinely active. A list that's continuously pruned toward active, consented recipients is how serious senders stay closer to the sub-0.1% complaint rates providers want to see.

Third, you suppress complaints and unsubscribes instantly and permanently. The moment someone opts out, they're gone from every future send — no accidental re-contact, which is one of the fastest ways to torch a reputation.

The obvious objection is, "but that gives me a much smaller list." Yes. And that's the point, not the cost. A smaller list of consented, engaged recipients beats a bloated cold list on every metric that actually matters: inbox placement, reply rate, conversion, and the long-term health of your domain. Ten thousand cold contacts that land you in spam are worth less than five hundred people who chose to hear from you. Cold email optimizes for the size of the send. The inbox optimizes for whether anyone wanted it. Build for the second one.

How we built this into GoodSender

This is the bet GoodSender is built on, so I'll be upfront that I'm describing our own product here. We took the permission-and-engagement playbook the biggest senders use and made it automatic, because expecting every small team to operate it by hand is unrealistic.

In practice that means two things run for you. The Permission Loop governs your custom marketing sends: GoodSender emails each recipient a short, high-reputation consent message with approve and reject buttons. An approval unlocks them so your sends go through; if they never approve, the message is simply never sent — so you physically cannot blast marketing mail at people who didn't agree. Transactional template emails — OTP codes, receipts, password resets, MFA prompts — sit outside the loop and send without a consent step, so the messages people are actively waiting for never get held up. The Engagement Check then tracks opens, clicks, bounces, and complaints continuously, re-confirms anyone who's gone quiet for six months, and suppresses unsubscribes and complaints across your whole workspace the instant they happen.

The Permission Loop gates only marketing email — transactional templates send without consent. An approval unlocks the recipient for every future marketing send, and the Engagement Check keeps the list healthy over time.

Because promotional sends are consent-gated and transactional sends are separated by purpose, complaint risk is reduced by design — which is also why we can offer the first 100,000 emails a month free and $1 per 100K after that. The economics only work when the list is clean, so the model rewards exactly the behavior the inbox already rewards.

None of this is a way to do cold email more sneakily. It's the opposite: a way to stop needing cold email at all, by making the permission-based alternative as fast and low-effort as the thing it replaces. If you've been feeling cold-email returns shrink quarter over quarter, that pressure isn't going to ease. The teams that move to earned, permission-based sending now are the ones most likely to still be landing in the inbox a year from now.

That's the whole idea behind what we're building. If it resonates, I'd love for you to give it a look — and tell me where I'm wrong.

How consent-gated email works (and why it's free)

Givi Pataridze — Thu, 21 May 2026 17:07:31 +0000

Email infrastructure pricing has barely moved in a decade. Mailgun, Postmark, SendGrid, Resend, all per-email, all clustered in a narrow band. The reason isn't bandwidth, and it isn't storage. It's abuse. Most of what an ESP charges you covers the cost of fighting the senders who shouldn't be on the platform in the first place.

Remove the abuse, and the price collapses.

That's the idea behind GoodSender's Permission Loop. Here's how it works, and why the economics finally let us run the first 100,000 emails a month at $0 and keep the curve flat after that.

The shared-pool problem

Every sender on a typical ESP sits in an IP reputation pool with everyone else on that tier. One bad actor blasts a purchased list, mailbox providers throttle the IPs, and your perfectly legitimate transactional mail starts landing in spam. You did nothing wrong. You're paying for the worst sender on your tier.

ESPs aren't oblivious to this. They staff anti-abuse teams, run heuristics, freeze accounts, review manually. That work is expensive. Per-email pricing is, in large part, a tax on the entire customer base to cover the cost of policing it.

The Permission Loop

GoodSender flips the order. Before you can send to a recipient, that recipient has to confirm they want your mail. The flow is simple.

You call POST /v1/emails/consent with the recipient's address.
We send a short, high-reputation consent email with approve and reject buttons. The link is signed, so the sender can't forge it.
Once the recipient clicks approve, that address is unlocked across your workspace. Sends go through instantly, from any API key in the workspace.
If they click reject, the address is suppressed. The original consent email stays live, so they can change their mind later and lift the suppression themselves.
If they never click, every send to that address is discarded inside GoodSender. Approve is a hard gate, enforced at the API.

Transactional templates (MFA Enrollment, New Device, Login, Order Completed, OTP Code) bypass the gate, because they're already implicitly requested by the recipient's own action. Transactional sends don't change consent state. They only fail if the recipient has explicitly clicked reject.

The Engagement Check keeps permission honest

A permission granted today isn't permission granted forever. We track opens, clicks, bounces, and complaints per recipient, aggregated into a live engagement score. Six months without engagement triggers an automatic re-permission ping. No response, the recipient drops off.

Unsubscribes and spam complaints are suppressed instantly across the entire workspace. No API key in that workspace can mail those addresses again, regardless of which engineer or which integration tries.

Permission is a snapshot. Engagement is the proof it's still real.

Why this lets us be free at the volumes most senders need

Permission-gated mail, by construction, has almost no abuse surface.

Purchased lists don't work, because the recipients haven't opted in to your sending domain.
Cold sequences don't work, because you can't email someone who hasn't confirmed.
There's no long-tail decay, because dormant recipients get pruned.

If abuse is structurally prevented, anti-abuse cost trends toward zero. Deliverability stays high without a fire-fighting team. The IP pool runs at unusually low complaint rates by design, and the marginal cost of a sent email, on infrastructure already serving the world's largest senders via Laneful, drops to roughly the cost of running the pipe.

That's why the first 100,000 emails a month, including consent requests, transactional templates, and marketing sends to consented recipients, cost $0. One bucket, no per-type splits, no tiered accounting. Beyond 100K it's $1 per additional 100,000. No subscriptions, no tiers, no credit card to start. The curve stays flat: roughly $9 a month at one million emails, $99 a month at ten million.¹

100,000 emails a month is about 3,300 sends a day, more than most SaaS apps and small or mid-sized businesses ever reach. For the majority of senders, the bill stays at $0 forever.

What this changes for senders

For indie developers, solo founders, and AI startups building agentic workflows, the time you used to spend on IP warmup, suppression lists, and deliverability tickets becomes a config call. Domain verification takes about five minutes. DKIM, SPF, and DMARC get configured automatically.

This isn't built for marketing teams running mass-broadcast campaigns. It isn't built for cold-outreach teams. It's built for makers and builders who want every recipient to be someone who actually asked.

Who this is for

The Permission Loop isn't a feature you can A/B test against an open rate. It's a constraint. It rules out workflows that other ESPs accommodate — purchased lists, cold sequences, lapsed-subscriber reactivation campaigns — and it rules them out at the API, not in the terms of service.

If your sending depends on those workflows, we're not the right tool, and we'll save you time by saying so now. If your sending is mostly transactional plus a newsletter people actually signed up for, the constraint is invisible and the bill is $0.

Pricing as of April 2026. ↩

Per-email pricing is a tax on bad senders. You're paying it.

Givi Pataridze — Wed, 20 May 2026 19:08:55 +0000

Open your last Mailgun invoice. The line says per email. It isn't.

A real email, the SMTP handshake, the DKIM signing, the bytes on the wire, costs a fraction of a fraction of a cent in 2026. AWS will rent you SES at roughly $0.10 per thousand. The marginal cost of a sent message has been rounding-error money for years.

So what are you paying for?

You're paying for the abuse team.

The line item nobody itemizes

Every general-purpose ESP runs an anti-abuse organization. Heuristic detection on outbound traffic. Manual review queues for new accounts. Suppression-list infrastructure. Relationships with mailbox providers and feedback loops with Gmail, Outlook, Yahoo. Lawyers responding to subpoenas. On-call engineers who page when an IP gets blocklisted at 3 a.m. because somebody else on the same shared pool decided to test a list they bought somewhere on the internet.

That work is real and necessary. It's also expensive, and it has to be paid for somehow. The way it gets paid for is the per-email price.

A useful exercise: line up the public list prices for 100,000 emails a month.

Mailgun: $75
Postmark: roughly $134
SendGrid: $34.95 (no free plan since May 2025)
Resend: $35, climbing to roughly $650 at one million emails
Amazon SES: roughly $10–12, if you bring your own AWS account, warm-up, and compliance posture

The infrastructure cost is roughly the same. The variance isn't bandwidth — it's how much abuse defense each platform absorbs.

The senders who pay the most

Here's the part that usually goes unsaid. Clean senders subsidize dirty ones.

If you run a SaaS app sending password resets and receipts, your complaint rate is probably under 0.01%. You generate roughly zero load on the abuse team. And yet you're paying the same per-message rate as the customer next door who's blasting a list they "warmed up" by importing it from a bankrupt LinkedIn scraper.

The pricing isn't malicious. It's structural. Shared infrastructure means shared cost. The platforms are doing the right thing operationally, pooling abuse defense across their customer base. But pooled costs are pooled by design. You can't pay less just because you behave better, because the platform has no mechanism to know you behave better until well after the bill has been printed.

Remove the abuse, remove the cost

The way out isn't a discount code. It's a different model. GoodSender's Permission Loop requires recipients to confirm before a sender can mail them, which means purchased lists and cold sequences don't function and the abuse surface is structurally near-zero.

Once that's true, the math changes:

No abuse team rotation to staff.
No suppression infrastructure to scale.
No cross-customer reputation contamination to clean up.

The marginal cost of an email drops back to what it actually is — pipe, plus a little metadata. That's why the first 100,000 emails a month cost $0, and why the curve stays flat at scale: roughly $9 a month at one million, $99 at ten million.¹ The model still doesn't tax you for someone else's behavior.

What to look at when you compare

If you're evaluating email APIs in 2026, don't compare per-email price first. Compare what's in the price — whether your reputation is pooled with the worst sender on your tier, whether the platform prevents abuse structurally or polices it with a team, and whether your per-email cost will track infrastructure or anti-abuse overhead as you scale.

The bet

Per-email pricing has held its shape for fifteen years because nobody had a structural reason to break it. The shared-pool model required pooled cost, pooled cost required per-message billing, and the abuse tax stayed embedded in the rate.

That's the part we think breaks next. Not because senders demanded it — they've been paying the tax without complaining for a long time — but because the model that makes the tax unnecessary is finally cheaper to run than the model that requires it.

Whether GoodSender is the platform that proves the point or somebody else is, the per-email rate as a category is on borrowed time.

Pricing current as of April 2026. ↩