The most-installed AI skill in the world is 22 words long

The most-installed skill on the planet is twenty-two words long. The man who wrote it has more GitHub stars than Anthropic’s official repo. He has also published the only serious theory of how to write a skill, and the biggest marketing pack in the world breaks every rule in it.

He writes at one seventh the length and has five times the stars. His longest skill is shorter than their median.

A million installs for twenty-two words

Matt Pocock’s grill-me is the most-installed non-bundled skill in the entire agent ecosystem. Roughly a million installs. Here it is, complete, nothing removed:

---
name: grill-me
description: A relentless interview to sharpen a plan or design.
disable-model-invocation: true
---

Call the Skill tool with "grilling".

Twenty-two words. One instruction. It doesn’t tell the model it’s an expert. It doesn’t have core principles. It doesn’t ask you five questions at the end. It points at the thing that does the work and gets out of the way.

What it actually does

You have a plan. A campaign, a launch, a page rebuild. Normally you describe it to an agent and the agent starts building, because agents are eager and agreeable and will happily construct the wrong thing beautifully.

Type grill-me instead and it refuses to build anything. It interviews you. It works out which decisions your plan depends on, asks the ones it can ask now, waits, and uses your answers to work out what to ask next. It keeps going until there’s nothing left to ask, and only then does it touch the work.

A good consultant asks why seven times. This is that, pointed at you, by the machine that’s about to do the work. It’s the discovery call nobody runs before prompting.

The interface is a batch of numbered questions, each arriving with the agent’s own recommended answer attached:

❓ Q1 - question title: question body, might be multiple
     paragraphs, including multiple choices

➡️ your recommended answer

---

❓ Q2 - question title: ...

➡️ your recommended answer

That recommended-answer line is what makes it usable rather than exhausting. You’re not sitting an exam. You’re reviewing a proposal, approving most of it and overruling the two things it got wrong, which takes a minute and surfaces the assumption that would have cost you a week.

Which makes one comparison unavoidable, because the marketing pack knows about this too. Here’s its complete treatment of the idea, from marketing-psychology:

Break problems down to basic truths and build solutions from
there. Instead of copying competitors, ask "why" repeatedly to
find root causes. Use the 5 Whys technique to tunnel down to
what really matters.

Name the technique, cite the number, move on to the next of seventy-two models. Forty words, and the reader is no more able to run an interrogation than before they read it.

Pocock named nothing. He wrote the procedure, gave it a stopping condition, and it became the most-installed skill on GitHub.

The thing doing the work, grilling, is 319 words and contains an actual algorithm:

Map this as a design tree: every decision branches
into the decisions that hang off it.

Work the tree in rounds. The frontier is every decision
whose prerequisites are already settled: the questions you
can ask _now_ without guessing at answers you haven't heard.
Ask the whole frontier in one round.

The session is done when the frontier is empty: every
branch of the design tree visited, nothing left silently
assumed.

Read that against a marketing skill’s “Core Principles: 1. Accuracy First. 2. Use JSON-LD. 3. Follow Google’s Guidelines.” One is a procedure a machine can execute and know when it has finished. The other is a poster in a break room.

Length is a confession

The obvious objection is that these are different jobs, and marketing genuinely needs more words than engineering. That objection is wrong, and the marketing pack proves it itself: its 897-word conversion skill and its 38,472-word marketing plan are in the same repo, and the short one is the bad one because it says nothing, not because it’s short.

Length in a skill isn’t thoroughness. It’s the sound of an author who hasn’t worked out what the agent should actually do.

Every one of those 45 persona openers is a writer buying time. “You are an expert in structured data and schema markup” is what you write when you haven’t decided what the procedure is. Pocock never writes it, in 37 attempts, because he always has.

He wrote down the rules. Nobody in marketing read them.

Buried in his repo is writing-for-agents, a genuine theory of how to write documents that agents consume. It’s the best thing published on the subject.

Pocock has never written about marketing and shows no sign of having read a marketing skill in his life. He was solving his own problem, for engineers. But because the rules are about how agents read documents rather than about any subject matter, they apply to every skill ever written.

• Context load. Always-loaded material spends tokens and attention every turn, whether or not it fires. A 4,450-word median is a permanent tax.
• Sprawl. “A document simply too long, even when every line is live and unique. Attention thins across the excess.”
• Completion criteria. Every step needs a condition that tells the agent it’s done. A vague bound invites premature completion, where the agent stops early because attention slips to being finished.
• One trigger per branch. “Synonyms that rename a single branch are one branch written twice; collapse them.”

That last rule is worth sitting with, because 44 of the 50 marketing skills break it deliberately. Their descriptions stuff in fifteen to thirty quoted synonyms so the skill triggers on as much as possible. There’s even a commit that boasts about it: “optimize all 32 skill descriptions for better triggering.”

By Pocock’s framework that isn’t clever. It’s one branch written twenty times, paying context load on every turn, to catch prompts the skill was never designed to handle well.

The rule that kills the questionnaire

Finding facts is your job, never the user’s. When a frontier question needs a fact from the environment, dispatch a sub-agent to find it; don’t ask the user for anything you could look up yourself. The decisions are the user’s: put each to them and wait.

Thirty-six of the fifty marketing skills end with a section called “Task-Specific Questions.” The schema skill’s list includes “Is there existing schema on the page?”

That’s a fact. It’s sitting in the page source. The agent could answer it in four seconds by fetching the URL. Instead the skill hands the question back to you, and you’re expected to be grateful for the consultation. Thirty-six skills do this. Pocock’s do it zero times, because he wrote down the rule that forbids it.

”Avoid spam tactics” is the worst line in marketing AI

Here’s the line, from the schema skill, sitting in a section titled Core Principles:

### 3. Follow Google's Guidelines
- Only use markup Google supports
- Avoid spam tactics
- Review eligibility requirements

I used to think that line was merely useless. Three wasted tokens telling a model to be good, when it was never going to commit schema fraud in the first place.

It’s worse than useless. It’s actively counterproductive, and Pocock explains exactly why:

Steering by prohibition drags the forbidden behaviour into context and makes it more available, not less. Don’t think of an elephant, and the elephant is all there is; the negation is a weak modifier the strongly-activated concept overruns, so the ban half-reads as an instruction to do the thing.

The instruction activates “spam tactics.” Then it applies a two-letter modifier and hopes. The skill has spent your context window putting spam in the room.

Toddlers, obviously

Anyone who’s been near a three-year-old already knows this. Say “don’t run” and you get running. Say “don’t touch that” and the hand moves.

That’s not a cute parallel. It’s the same mechanism, and the developmental literature has measured it. Two-year-olds disregard the negative element entirely and process negative sentences as positive. Three-year-olds are still at chance. Children don’t reliably resolve negation until around four. Both children and adults activate the to-be-negated information while processing the negation, and the recommendation from the research is blunt: avoid negation in instructions, say the positive instead. “Do touch the cold part” beats “don’t touch the stove.”

Two literatures, one mechanism. The concept fires first, the negation arrives late and weak, and it fails to suppress. Both arrive at the identical fix:

Prompt the positive. State the behaviour you want, so the one you don’t want is never spoken.

“Avoid spam tactics” should be “Mark up only content visible on the page.” Same intent, no elephant. It takes ten seconds to fix, and in eight months nobody has, across a pack with 4.6 million installs.

Steal these four rules

You don’t have to write TypeScript skills to use any of this. Four rules, taken straight from the most successful skill author on GitHub, that would fix most of what’s wrong with marketing skills tomorrow.

1. Cut the persona. If your skill opens by telling the model it’s an expert, you haven’t decided what it does yet. Go decide, then write that instead.
2. Give every step a stopping condition. “The session is done when the frontier is empty” is a bound a machine can check. “Analyse the page across these dimensions” is not.
3. Never ask for a fact you can fetch. Decisions belong to the human. Facts belong to the agent. Every questionnaire at the end of a skill is that skill outsourcing its own job to you.
4. Say the positive. Every prohibition in your skill is an elephant. Replace it with the behaviour you actually want.

The marketing skills category has this exactly backwards. It competes on coverage and length, because those are the things you can see from the outside, the same way it competes on install counts and star badges. Pocock competes on whether the thing works, which is why his median skill is 595 words and why a twenty-two word file is the most installed in the world.

The whole marketing pack could be rewritten at a fifth of the length and would be better for it. Somebody should do that. It’d take a fortnight.