跪拜 Guibai
← All articles
Frontend · AI Programming · GitHub

grill-me: An AI Skill That Interrogates Your Vague Ideas Before You Write a Single Line of Code

By JacksonChen ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most AI coding failures trace back to unclear requirements, not poor code generation. grill-me enforces the engineering discipline of clarifying what to build before building it, turning a fuzzy one-liner into a boundary-defined spec through structured interrogation.

Summary

Misalignment — when AI generates code that doesn't match what you actually wanted — is the most common failure mode in AI-assisted programming. grill-me, part of Matt Pocock's 176k-star skills repository, tackles this by acting as a relentless interviewer before any code is written. It walks down the decision tree branch by branch, asking one question at a time with a recommended answer, and refuses to proceed until the developer confirms a shared understanding.

The skill deliberately separates orchestration from execution: grill-me is user-invoked, meaning the human decides when to trigger it, while the AI handles the questioning discipline. It only asks about decisions the developer must make, looking up discoverable facts from the codebase itself.

grill-me sits at the entry point of a full pipeline — alignment, specification, ticket decomposition, TDD-driven implementation, and dual-axis code review — but the recommendation is to start with grill-me alone. The real value isn't the AI making decisions for you; it's that the questioning process forces you to think through your own requirements to a depth you wouldn't reach on your own.

Takeaways
Misalignment is the most common AI Agent programming failure mode, not code generation quality.
grill-me asks one question at a time with a recommended answer, waiting for the developer's response before continuing.
The skill looks up discoverable facts from the codebase itself and only asks about decisions the developer must make.
It refuses to write any code until the developer confirms a shared understanding has been reached.
grill-me is user-invoked by design — the human decides when to trigger it, the AI handles the questioning discipline.
It is the entry point of a pipeline: /grill-me → /to-spec → /to-tickets → /implement → /code-review.
Most users end up accepting the AI's recommended answers, but the process still deepens their understanding of the requirement.
Conclusions

grill-me inverts the typical AI coding dynamic: instead of the developer prompting the AI, the AI prompts the developer until the idea is sharp enough to implement.

The skill's design principle — humans orchestrate, agents execute — draws a clear line that many AI coding tools blur, and it may reduce the thrash of iterative re-prompting.

Accepting every AI recommendation still produces value because the structured questioning itself builds the developer's mental model of the requirement, which is the real bottleneck in most failed AI coding sessions.

Concepts & terms
Misalignment
In AI-assisted programming, the failure mode where generated code does not match the developer's actual intent, typically because the requirements were never clearly specified before generation began.
User-invoked vs. Model-invoked skills
A design distinction in Matt Pocock's skills system: user-invoked skills are triggered manually by the developer for orchestration decisions (when to do what), while model-invoked skills are triggered automatically by the AI for execution discipline (how to do it).
Tracer-bullet vertical slices
A development approach where each ticket or work item cuts through all layers of the stack (schema, API, UI, tests) to deliver a thin but complete slice of functionality, rather than building layer by layer.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗