At 1,005,653 installs, the second most-installed agent skill in our catalog is seven lines long. Not seven hundred. Seven. As of August 30, 2026, Matt Pocock's grill-me skill trails only the find-skills skill from Vercel Labs, which holds 3,170,600. The installed file writes no code and contains no method. Its one working instruction hands off to a sibling.
Disclosure: we run Skillselion, an independent catalog refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs. We read both files line by line at commit 885e2ca of mattpocock/skills on August 20, 2026, as part of our deep dive series. Every quote below is from that commit.
What does the seven-line grill-me skill actually do?
Call the Skill tool with "grilling".
Source: SKILL.md, full file at the grill-me folder in mattpocock/skills
That is routing, not laziness. The repo keeps one method skill named grilling and points two trigger skills at it: grill-me for plans, and the grill-with-docs skill for architecture decisions against a live codebase. grill-with-docs is itself at 856,939 installs, third in the catalog. The dependency matters when you install: seven lines have nothing to fall back on, so grill-me does nothing unless the grilling skill is present too.
Why can the agent not start grill-me by itself?
disable-model-invocation: true
Your agent cannot fire grill-me on its own. You say "grill me", or nothing happens. For a skill whose entire output is questions aimed at you, that is the right lock. The interesting part is where the lock is not: grilling, the method underneath, carries no such flag, and its description invites the agent to reach for it when you seem to want your thinking stress-tested. The trigger phrase belongs to you. The method keeps its own judgment.
How does grill-me structure its questions?
The method models your plan as a tree where every decision branches into the decisions hanging off it. The set of questions whose prerequisites are already settled is called the frontier, and the agent asks the whole frontier in one numbered round:
❓ Q1 - <question title>: <question body>
➡️ <your recommended answer>
Each question ships with the agent's own recommendation, so you can answer "yes to your suggestion" instead of writing an essay. A question that depends on another still-open question belongs to a later round, explicitly. Every answered round reshapes the tree.
How does grill-me avoid asking you to run ls?
Finding facts is your job, never the user's.
Source: grilling SKILL.md
Anything the agent could look up in the environment, it dispatches a sub-agent to find instead of asking you. A running lookup counts as an unsettled prerequisite, so only the questions downstream of it wait while the rest of the frontier gets asked. You never get asked to run ls.
The exit condition is testable rather than vibes based: the session ends when the frontier is empty, and even then the agent must not act until you confirm you have reached shared understanding.
What changes once grill-me is installed?
| You say | Without grill-me | With grill-me |
|---|---|---|
| "Grill me on this plan" | The agent agrees and starts building | Numbered question rounds until no branch is silently assumed |
| "We should add a Postgres cache here" | "Great idea", then an implementation | The frontier fills with invalidation, the local dev story, and who owns the schema |
npx skills add https://github.com/mattpocock/skills --skill grill-me
Install notes and the trigger vocabulary are in the grill-me guide.
Why do mattpocock/skills entries hold seven of the top ten?
Look at the top ten by installs on August 30, 2026 and seven slots belong to skills from the mattpocock/skills repository. The tell is the setup-matt-pocock-skills skill at 734,383 installs in eighth place. That skill does nothing but scaffold the per-repo configuration the other skills assume: issue locations, triage labels, domain-doc layout. On its own it is useless. The likeliest read is that people install the author's stack rather than picking individual files. The tdd skill at 797,033 and the improve-codebase-architecture skill at 824,417 ride the same wave.
The other pattern in that list: skills that stop the agent outrank skills that make it produce. The same shape shows up across the August 2026 agent economy census. If you want to compare planning skills against each other, the docs and planning hub ranks that category, and the full skills catalog covers the rest.
Install counts are live as of August 30, 2026. Quotes are from commit 885e2ca, verified August 20, 2026.
Top comments (0)