DEV Community: sazerfenor

Why my AI photo effects site has no prompt box

sazerfenor — Sun, 15 Nov 2026 05:12:07 +0000

I run a small site that turns one photo into a finished AI effect: a Christmas card, a figurine, a restored family print. The decision that shaped everything else was leaving out the prompt box. This is a build log of that choice and the product decisions that followed from it.

The problem with a blank text field

Most image tools open with an empty text area. For people who generate images every day, that is freedom. For someone who just wants a Christmas card from a family photo, it is homework: guess what to type, wait, rewrite, wait again. On a paid product each retry can cost money.

So I chose presets instead of a free prompt, because a casual user would rather click a style they can see than learn to describe one. Each effect page shows sample results and a short list of styles. The Christmas card page, for example, offers five styles, and the figurine page offers three: boxed classic, anime and clay toy. You upload one photo, click a style, press Generate.

There is a cost, and the pages say it: free-text prompts are not supported, because the styles are fixed. If you want full control, this is the wrong tool. As a nod to people who like prompts, each effect page also lists a copyable prompt per style for pasting into a chat image tool, with a note that I have not tested those elsewhere.

For a developer, presets turn an open-ended problem into a small, enumerable one: five styles are five things to test, five sample images to show, and one price per run.

One photo in

I chose one photo per run because multi-photo input adds more decisions for the user. The pages are honest about the edges: merging several separate photos is not what it does, group photos with many faces give weaker results, and an exact likeness is not guaranteed every time. Saying so up front saves people wasted credits.

Input limits

Accepted files are JPG, PNG and WebP, up to 8 MB. I kept that list short on purpose: it fits in one line under the upload box, and it is shown on every effect page, so nobody discovers the limit after a failed attempt.

For the old-photo restoration page I also added scanning advice: a flatbed scan works best, or lay the print flat in daylight with no flash. Input quality is the biggest lever on output quality, and that advice is cheaper to give than to fix afterwards.

Pricing in credits, not time

Every image costs 10 credits, whatever the effect. One balance works across all effects. I chose a flat credit cost because users can predict their spend before they click. Packs are one-time: $4.99 for 50 credits, $9.99 for 120, $19.99 for 300, so one image costs between $0.67 and $1.00 depending on the pack. Pack credits stay valid for 365 days. Monthly plans start at $9.90 for 500 credits, and the larger plan brings a single image down to about $0.13.

New accounts get 10 free credits while the day's new-account gifts last, enough for one image. You sign in before generating.

Failures should not cost the user

AI image generation fails sometimes. A render can error or time out. I chose to return credits automatically in those cases, because a user who loses credits to a failure has no reason to trust the product again. The refund policy says that if a generation fails, does not finish in time, or is declined by the image model, the credits come back to the balance with no ticket needed.

Treat refund-on-failure as part of the product, not a support gesture. Credits spent on a result that did finish are a different case and are not refundable, though an unused pack can be refunded within 7 days.

Short retention for uploads

Photos of people and pets are personal. I chose to delete uploads 24 hours after the generation finishes, whether it succeeded or failed, because I do not need the original after the result exists. Results are kept for 30 days so people can download again from their creations list, then removed. Files are shown only through signed links that expire.

I also tried to be straight about the limit of that promise. The photo has to be sent to an outside image provider to make the result, and my deletion applies to my own copies, not theirs. The privacy page says so. If you build something similar, write that sentence down early, because users will ask.

Around 37 seconds

Each result takes about 37 seconds. I put the number on every page instead of saying "fast", because a number sets an expectation people can check.

What the site is

Effect Booth is an independent service. It is not affiliated with or endorsed by the providers of the AI models it uses. It applies AI photo effects to your own photos: upload one photo of a person or pet, pick a preset style, download the result. Current effects include AI Christmas card, Photo to figurine and Restore old photos; the full list is on the home page. You pay by card through Stripe Checkout, you sign in first, and a result is a digital image only: nothing is printed or shipped.

Takeaways

For casual users, fixed styles remove the hardest step. Keep the prompt box only if your users want it.
Say what the tool does badly. It reduces refunds and bad first impressions.
Keep input limits short and visible next to the upload control.
Price per image in a single unit so the cost is knowable before the click.
Refund failed runs automatically, and write down what happens to uploads and how long.

If you have built something similar and chose differently, I would like to hear why.

What one chat reply costs across 330 AI models (Nov 2026)

sazerfenor — Sun, 15 Nov 2026 05:11:24 +0000

When people compare LLMs, they usually compare per-million-token prices. Those numbers are accurate, but they are awkward to reason about when what you actually do is send a message and read an answer. I wanted a more concrete question: what does one ordinary chat reply cost, and how much does that change from model to model?

I built a small site that lists models with prices and an estimated cost per reply, so I had the data on hand. This post walks through how the math works and what the spread looks like as of mid-November 2026.

From per-1M prices to a per-reply cost

Every model is priced in dollars per 1 million input tokens and per 1 million output tokens. A single reply costs:

cost = (input_tokens / 1,000,000) * input_price
     + (output_tokens / 1,000,000) * output_price

Two details matter in practice:

Input includes history. In a chat, each new message sends your text plus the earlier messages, so input grows as a conversation goes on.
Output usually costs several times more than input. On the models below, output is 2.5x to 5x the input price. Long answers are where the money goes.

For the estimates on my site I use a fixed "typical reply": about 300 words in and 500 words out. Real replies are billed by their own token counts, so a short question with a short answer costs less and a long conversation costs more. The typical reply is just a way to compare models on equal footing.

The spread, in a table

These are prices from OpenRouter's public model list, synced on November 14, 2026. The last column is the site's estimate of credits per typical reply, where one credit is about $0.01 at pack prices.

Model	Input / 1M	Output / 1M	Per reply
LiquidAI LFM3-1.7B	$0.08	$0.20	about 1 credit
OpenAI GPT-6 Luna	$0.10	$0.50	about 1 credit
Thinking Machines Inkling 2 Small	$0.30	$1.00	about 1 credit
Baidu ERNIE-Light-DT-1.0	$1.13	$3.94	about 1 credit
Google Gemini 4 Hydrogen	$1.50	$7.50	about 2 credits
Google Gemini 4 Carbon	$2.00	$10	about 3 credits
Anthropic Claude Opus 5.5	$4.00	$20	about 5 credits
Anthropic Claude Fable 5.2	$8.00	$40	about 9 credits

A few things stand out.

The spread is large. Output on Claude Fable 5.2 is $40 per million tokens, against $0.50 for GPT-6 Luna. That is an 80x difference in output price. In reply terms it is about 9 credits versus about 1, roughly nine cents versus one cent.

There is a floor. The ERNIE row costs more per token than Luna or Inkling, yet it still shows about 1 credit. That is because each reply costs at least 1 credit, so below a certain price the per-token differences stop showing up on the bill. For small models, the question is no longer "which is cheapest" but "which is good enough".

Doubling the price roughly doubles the reply cost. Opus 5.5 ($4 / $20) is half of Fable 5.2 ($8 / $40), and the estimate goes from about 9 credits to about 5. Anthropic's own guidance, as quoted on the Fable 5.2 vs Opus 5.5 comparison, is to start most work with Opus 5.5 and move up to Fable 5.2 for more demanding reasoning and long-horizon agent work, or when your own evaluations at higher effort still fall short. That is a sensible rule for any pair of models: start with the cheaper one, and move up when you can point to a failure.

Prices move. Gemini 4 Carbon is listed at $2 / $10 until November 30, 2026 and $4 / $20 from December 1, per Google's pricing page, as summarized on the DeepSeek V4.1 Flash vs Gemini 4 Carbon page. Whether OpenRouter follows that change was not confirmed when I last checked. If you budget from a table like the one above, put a date on it.

When a cheaper model is good enough

I have no benchmark scores to offer, and I would rather not invent any. As a rule of thumb, the cheap end is often fine for:

Rewriting, summarizing, or reformatting text you already have
Short factual questions and quick explanations
Drafting boilerplate such as email subjects, commit messages, or docstrings
First-pass code questions where you will check the answer anyway

The pricier models earn their cost on work where a wrong answer is expensive for you: long multi-step reasoning, large refactors, or agent loops that run for a while. A practical habit: send the same prompt to a cheap and an expensive model. If you cannot tell the difference, you have your answer.

Conversation length is another lever: history is re-sent as input, so a long thread on an expensive model gets costly. A fresh chat per topic is the cheapest optimization there is.

The site

The site is AI Model Buffet. You pick from 331 text models and chat with all of them from one credit balance. Each reply is charged by its real token cost, with the service fee included and a minimum of 1 credit. A new account gets 5 free credits once, valid for 7 days, and paid plans start at $9.90 per month (1,200 credits a month). There are also one-time credit packs.

The conversation is kept in your browser; the server stores only the model, credits, token counts and cost of each reply, not the text. It is an independent service and is not affiliated with or endorsed by OpenRouter or any model maker. Replies are AI-generated and can be wrong, like any chatbot.

Takeaways

Per-reply cost is input tokens times the input price plus output tokens times the output price, and output usually dominates.
Between the cheapest and priciest listed models, the gap is around 80x on output price, and about 9x in typical credits per reply.
Many small models hit a billing floor, so compare quality, not just price, once you are down there.

If you have a workload where a cheap model surprised you, or an expensive one clearly paid for itself, I would like to hear about it in the comments.

How to install SKILL.md skills in Google Antigravity (and 5 worth trying first)

sazerfenor — Sat, 14 Nov 2026 00:50:46 +0000

A skill is just a folder with a SKILL.md file in it. Put that folder where your agent looks for skills, start a new conversation, and the agent loads it when your request matches the skill's description.

Where the folder goes

Agent	Project folder	Global folder
Google Antigravity	`<project>/.agents/skills/<skill-name>/`	`~/.gemini/config/skills/<skill-name>/`
Claude Code	`<project>/.claude/skills/<skill-name>/`	`~/.claude/skills/<skill-name>/`
OpenAI Codex	`<project>/.agents/skills/<skill-name>/`	`~/.agents/skills/<skill-name>/`

Each <skill-name> folder must contain SKILL.md directly, plus any scripts or reference files that ship with it. Antigravity still reads the older .agent/skills/ (singular) path, but new installs should use .agents/skills/. If a skill doesn't load, check that SKILL.md isn't nested one folder too deep. That's the most common mistake. The full install guide covers the other fixes.

Five skills worth installing first

Emil Kowalski's design and motion skill: teaches your agent purposeful UI motion, including easing, durations, springs and accessible interaction states. Good if your agent's animations feel random.
Matt Pocock's Teach skill: turns /teach into a learning workspace with a mission, lessons and records, so you can pick a topic back up in the next session.
Superpowers: Jesse Vincent's development-methodology suite (14 skills) that routes each task through the right process before the agent starts coding.
The Banana skill for Claude: Agrici Daniel's plugin for generating and editing images with Gemini from a visual brief.
Caveman: makes replies terse without touching code, paths, numbers or quoted errors. Useful when you're tired of scrolling.

All five come from their original authors' public repos, and you can install them from there. I collected them, with install-ready downloads, at Antigravity Skills. Disclosure: that's my site.

Which skill do you install in every new project? I'm looking for the next ones to try.