DEV Community: Paul Irolla

Anthropic admitted that Fable burned through their €200/month Claude Code plan in 1 hour instead of 5.

Paul Irolla — Tue, 07 Jul 2026 11:30:24 +0000

The community responded with an explosion of tools. Here are all the optimization techniques I've found. Ranked from C-Tier (meme) to S-Tier (serious engineering).

⭐ C-Tier: Use /compact when you're at 60% of the conversation's max context. 60% is the sweet spot. Context tokens aren't all equal, and when you're near the limit, agents start drifting and follow instructions less reliably. Why meme-tier? Because nobody's going to do it manually. (It's done automatically by the S-Tier tool.)

⭐⭐ B-Tier: Compress prompts before they reach the model. In any prompt, roughly half the words carry no real information, and agents are perfectly capable of filling in the blanks. That's why there's a whole category of specialized models for prompt compression with minimal quality loss: LLMLingua or Bear-2, for example. Why B-Tier? It's only really suitable for deterministic workflows (like n8n or similar), not for highly variable agentic environments. But it's excellent in those specific cases.

⭐⭐ B-Tier: Always tell Claude to launch its exploration sub-agents with Haiku instead of Fable, Opus, or Sonnet. This can be up to 92% cheaper on exploratory searches. Haiku is great at digging, not at thinking.

⭐⭐⭐ A-Tier: /effort default to xhigh. All the time. But 80% of what you ask Claude doesn't actually need that level of token waste. Switch to medium or high for most tasks — it cuts the amount of thinking required by 2x or 4x to reach the result.

⭐⭐⭐⭐ S-Tier: Agent harness compaction tool. Why S-Tier? It's plug-and-play and the gains are immediate, no matter what you're doing with your agent. Tokenade, for example, automatically optimizes: the commands the model uses, file re-reads, code/web search, MCPs, the order of tool calls, etc. It currently ranks #1 on THOL (Token-Harness Optimizer Leaderboard), with -27% token cost on 200k+ token conversations and -40% on 400k+ token conversations (Opus & Fable).

Do you use any of these techniques ?

Tokenade : https://tokenade.net/
THOL : https://pi-infected.github.io/token-harness-optimizer-leaderboard/
LLMLingua : https://huggingface.co/spaces/microsoft/llmlingua-2
Bear-2 : https://thetokencompany.com/

Position-1 CTR isn't 30% anymore — what AI Overviews actually do to your clicks (+ a free calculator)

Paul Irolla — Sat, 16 May 2026 08:48:29 +0000

For years, the SEO planning template was the same spreadsheet trick: pull the keyword volume, multiply by an assumed CTR (the famous "31% for position 1"), and call that your traffic projection. In 2026, that math is broken.

What changed:

AI Overviews sit above the organic results on most informational queries. The user often gets their answer in the box and never scrolls.
Featured snippets still steal the click on definitional / how-to queries.
People Also Ask clusters expand inline and burn 6-8 vertical inches of SERP real estate.
Sponsored carousels on commercial intent queries push the first organic result below the fold on mobile.

The result: position 1 organic CTR for an informational keyword in 2026 is closer to 8-15% than to the legacy 30%. For commercial-intent keywords with a sponsored carousel and an AI Overview, position 1 organic CTR can drop to 3-6%.

If you're still planning content with a 30% assumption, your traffic forecasts are off by a factor of 3-10x. Which means the keyword you're killing for "low projected upside" might actually be the one worth shipping, and the keyword you're prioritizing for "30% × 12k searches" is going to disappoint.

What I changed in my own planning

I split CTR assumptions into 4 SERP archetypes:

SERP type	Position 1 CTR (2026)	Position 2-3 CTR
Plain organic (rare)	28-35%	12-18%
Organic + PAA only	18-24%	8-12%
Organic + AI Overview	8-15%	3-7%
Organic + AI Overview + sponsored carousel	3-6%	1-3%

Before committing to a topic, I run the SERP through a quick check: which archetype is it? Then I plug the right CTR into the volume math. That single change cut the gap between my projected and observed traffic by ~70%.

A free thing for the lazy version of you

I built a free in-browser calculator that does this for you — paste the keyword + chosen position + SERP archetype, get a realistic click projection: the-seo-autopilot.com/en/free-tools/serp-ctr-calculator. No signup, all client-side, no email gate. Use it when you're triaging a 200-keyword list and don't want to manually classify every SERP.

There's also a related one I use a lot — a keyword opportunity calculator that combines this CTR adjustment with difficulty / volume / business value to spit out a single ranked score per keyword. Useful when you have a backlog of 50 ideas and need to pick the next 5.

TL;DR

The "30% for position 1" rule of thumb is a 2018 number. Don't plan with it in 2026.
Classify each keyword's SERP into one of 4 archetypes before assigning a CTR.
AI Overviews + sponsored carousels can drop your effective CTR by 5-10x.
Recalibrate. Your "low-volume but high-intent" keywords are probably more valuable than your "high-volume informational" ones now.

Plan your content with the SERP you actually have, not the SERP you wish you had.

What is /llms.txt? — the AI-crawler hint file (+ the free generator I built for it)

Paul Irolla — Fri, 15 May 2026 17:38:46 +0000

Last year a small RFC started making the rounds: /llms.txt. It's a single-file convention that tells AI assistants (ChatGPT, Claude, Perplexity, the Google AI Overview, Copilot…) which content on your site to surface and how to summarize it. Think robots.txt, but for meaning instead of crawl rules.

I've been running an autonomous SEO agent on a SaaS site, and /llms.txt showed up as the cheapest-yield optimization I could ship: a single static file, written once, that materially changes how an LLM frames your brand when a user asks about your space.

This post is what I wish I'd read before writing my first one.

What `/llms.txt` actually contains

The format is intentionally minimal — a markdown file at your domain root:

# Acme — autonomous test runners for SaaS

> One-line description of the project, exactly the way you'd want an
> LLM to re-quote it.

## Docs
- [Getting started](https://acme.dev/docs/start.md)
- [API reference](https://acme.dev/docs/api.md)

## Examples
- [End-to-end test for a Stripe checkout](https://acme.dev/examples/stripe.md)

## Optional
- [Pricing](https://acme.dev/pricing.md)

Three things matter:

The summary line is the brand-positioning weapon. If a user asks "what is Acme?", the LLM is statistically going to lift this line. Write it like the headline you'd put on a billboard, not like a feature list.
Sections are signals. "Docs" / "Examples" / "API" tell the LLM what surfaces are worth pulling into a citation. The order matters — what you list first gets weighted higher.
Optional vs required. The ## Optional section is content you'd be happy to see referenced, but don't need surfaced. Pricing, blog posts, careers — that kind of thing.

What it's NOT

It's NOT a robots.txt. AI crawlers will still hit your site without one.
It's NOT a sitemap. It doesn't enumerate every URL — it picks the important ones and tells the LLM how to frame them.
It's NOT machine-validated. There's no "llms.txt parser" that errors out on bad syntax. The LLM does its best with whatever you write. So write it well.

How to validate yours

I built a free in-browser generator that does the formatting + a sanity check on the structure: the-seo-autopilot.com/en/free-tools/llms-txt-generator. No signup — it's all client-side. Paste your site description, your section URLs, hit generate, copy the markdown into your site root.

If you want to see one in production, mine is at the-seo-autopilot.com/llms.txt. It's also indexed in SecretiveShell/Awesome-llms-txt — a curated list of all the public /llms.txt files in the wild, which is where I'd recommend looking before writing your own.

Does it actually move the needle?

Honest answer: too early to say with high confidence in 2026, but the signal is positive. Sites with a well-written /llms.txt get cited more often in ChatGPT and Perplexity answers in the niche I track (programmatic SEO + indie SaaS). The cost is roughly an hour, the upside is "free distribution surface that didn't exist 18 months ago."

Worth shipping.

I'm running an autonomous AI agent that publishes a fact-checked SEO article every day — here's the stack

Paul Irolla — Fri, 15 May 2026 16:18:46 +0000

For the past 60 days I've been running an experiment: an AI agent owns the entire SEO pipeline of one of my SaaS sites. Keyword research, drafting, fact-checking, internal linking, schema markup, scheduling — the agent does it all. Zero human touch on the content side.

Results so far: 0 → 400K+ Google impressions in 60 days. 46 articles live, 330+ scheduled. No backlinks bought, no paid promotion.

I'm sharing the stack because every "AI SEO" tool on the market today is either a glorified ChatGPT wrapper or charges $1k+/mo for marginal automation. The agent below runs on a small VPS in a Python loop and costs me less than $100/month in OpenRouter API credits.

The pipeline

Keyword pipeline — a daily cron pulls the SERPs of ~300 seed queries via DataForSEO, scores by intent × difficulty, picks the next batch.
Drafting — a frontier model writes a first draft against a structured outline. Cost: ~$0.10 per article.
Fact-check pass — a smaller, cheaper model re-reads the draft with a "find every factual claim and verify it has a credible source" prompt. Anything dubious gets flagged.
On-page — JSON-LD schema, internal links to existing articles via cosine similarity on embeddings, title/meta optimization.
Scheduling — one article goes live per day at a fixed time. The cadence matters more than the volume of any single drop.
Monitoring — a tiny script pings Google Search Console weekly and surfaces underperformers for the agent to refresh.

What I've learned

Quality from AI agents is fundamentally a context problem. The agent gets the keyword + the SERP context + the existing internal-link graph + a brand voice guide. With those inputs, modern LLMs write better than most freelancers I've hired for the same money.
The fact-check pass is the unlock. Without it, AI articles confidently hallucinate stats. With it, they're often more reliable than the median human content for the same niche.
Publishing daily compounds. Google indexes a healthy domain quickly when the cadence is regular. The 60-day curve looked basically flat for the first three weeks, then bent up.
Most "AI tool" directories will reject you if you aren't actually a tool. A playbook is not a tool. A free in-browser /llms.txt generator IS a tool. Frame accordingly when you submit.

Tooling

The site ships a free /llms.txt generator plus 12 other one-shot SEO tools (JSON-LD schema generator, SERP CTR calculator, hreflang tag generator, etc.) at the-seo-autopilot.com/en/free-tools — no signup, all client-side.

The full architecture writeup, prompts, and scheduler code are documented at the-seo-autopilot.com for indie SaaS founders who want to copy the setup. Happy to dig into specific bits in the comments — particularly the fact-check prompt, which is the part I had to iterate on the most.

Dominate LinkedIn in just 5 minutes a day

Paul Irolla — Sun, 10 Aug 2025 13:10:19 +0000

That’s not clickbait.
That’s exactly what Meet Lea was built for.

Here’s the problem:

You know you should post regularly… but blank page syndrome is real.

You want to engage with your audience… but comment replies eat hours.

You’d love to track what works… but analytics feel like a chore.

Lea solves all of it.

✅ Create posts that actually perform — Turn any idea, article, or file into a ready-to-publish LinkedIn post in seconds.
✅ Engage without the burnout — Lea suggests smart, human-like replies to every comment.
✅ Know exactly what to post next — Intelligent content analysis tells you what your audience truly cares about.
✅ Grow with clarity — Engagement tracking, growth insights, and post ideas… all in one dashboard.

Who’s already using it?
Freelancers. Agencies. Marketing managers. Community builders.
Over 500 people are on board — and in beta, Lea has already generated:

107K sources of content
17K post ideas
3K full posts
19K comments

This is your unfair LinkedIn advantage.
Access is limited. The waitlist is filling fast.

🎁 Join now and get:

Priority access

7 days free

Early adopter perks

🚀 Stop overthinking. Start dominating.
👉 Join the waitlist today.

https://meet-lea.com