Matt Pocock's fastest-growing skill this week is grill-me (GitHub), a skill whose whole job is to interrogate you about your plan before Claude Code writes a single line. It sits at 654,849 all-time installs, with 9,611 added in the last week, leading every other skill in the most popular skill repo we track.
It's not alone. The whole "grilling" family is climbing:
- grill-me (GitHub): 654,849 installs, +9,611 this week. Runs a grilling session that challenges your feature idea until the spec stops being vibes.
- grill-with-docs (GitHub): 554,895 installs, +8,074. Same interview, but it stress-tests your plan against your project's own documented decisions and updates the docs as answers land.
- grilling (GitHub): 266,286 installs, +7,827. The underlying interrogation engine the other two invoke.
That's roughly 25,500 installs in seven days for one idea: make the agent push back.
Anthropic moved the same week
On July 22, Claude Code v2.1.218 changed /code-review to run as a background subagent, so review work no longer eats your conversation. Three days earlier, v2.1.215 stopped Claude from running /verify and /code-review on its own, handing the trigger back to the developer.
Read those together with the install data and a picture forms. The tooling is converging on two bookends around code generation:
- Before code: interrogation. grill-me forces the spec conversation you were going to skip.
- After code: review. code-review (GitHub) is at 174,051 installs, +7,191 this week, and tdd (GitHub), the red-green loop that makes the agent prove its work, holds 517,422 installs, +7,296.
The middle part, writing the code, is the part with no breakout skill. Generation is table stakes now; the growth is at the bookends.
Why developers want an AI that argues
Agents made wrong code cheap to produce in bulk. A misunderstood requirement used to cost you an afternoon; now it costs you a 2,000-line diff that compiles, passes a smoke test, and is wrong. The expensive failure moved upstream, from writing code to deciding what to build.
The install numbers say developers have noticed. Last week the fastest risers in our search data were restraint skills that make Claude quieter and terser. This week it's skills that make Claude harder to satisfy. Same trend, next stage: first developers muzzled the agent, now they're hiring it as the skeptic in the room.
There's something honest about a market where the top of the charts is an AI that says "convince me first."
The full ranked list, refreshed daily from skills.sh, GitHub and MCP registries and ranked by real installs, is at skillselion.com.
I run Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs. Install numbers above are from the live catalog, pulled today.
Top comments (0)