Stop Letting Claude Code Guess What You Want: The grill-me Skill Claude Code Workflow
The grill-me skill Claude Code workflow stops AI from coding before it understands the problem. Here's how it works and how I'm applying it to my own SKILL.md builds.
Last week one of my Diploma IT students asked Claude Code to “build a login system” for his FYP. Ten minutes later he had Firebase auth, a Stripe billing hook, and three new database tables — none of which his supervisor asked for. That’s the failure mode the grill-me skill Claude Code workflow was built to kill: an agent that starts building before anyone agrees on what “done” looks like.
Seven million downloads for one skill file with four rules is a signal worth paying attention to. Here’s what it does, why the mistake it fixes is so common, and how I’m folding it into my own skill-building work at SUC.
The Common Mistake: Letting AI Plan Before You’ve Agreed on the Problem
You type one sentence. Claude Code is already three steps into a schema design. It looks efficient — until you realise nobody decided why you’re building the thing in the first place. Membership system for retention? Revenue? Screening? Nobody asked, so the agent picked for you.
This is the pattern I see most with my own students and HRDC workshop participants: they treat the AI’s speed as a proxy for correctness. Fast and wrong is still wrong, and with agentic coding tools the rework compounds — every revision on a bad assumption makes the codebase harder to unwind.
What grill-me Actually Does
Matt Pocock’s /grill-me is a slash-command skill with almost no cleverness in it. Four rules:
Ask only one question at a time.
Include a recommended answer with each question.
Never ask the user for information the AI can find itself.
Don’t start implementation until both sides agree on what needs to be done.
That’s the entire mechanism. It’s a brake, not a planner — its whole job is to stop Claude Code from rushing into “how should we build it” before anyone has answered “why are we building it.”
The Full Workflow: grill-me → to-spec → to-tickets
The brake alone isn’t the value — it’s the sequence that follows it:
/grill-me— clarify the actual problem through structured questions./to-spec— turn the agreed answers into a specification. Crucially, the spec describes constraints, not old code — copy-pasting existing implementation details into a spec means the next task inherits outdated assumptions./to-tickets— break the spec into tasks split vertically by user function (”a member logs in and sees benefits”), not horizontally by technical layer (”build the DB, then the API, then the frontend”). Vertical slices mean you find out immediately if an assumption was wrong, instead of discovering it at final integration.TDD implementation — write code while locking the success criteria so the agent can’t quietly redefine “done.”
Review and architecture checks — catch the codebase fragmenting into disconnected pieces.
Real Example: Applying This to My Own SKILL.md Files
I write a lot of SKILL.md files — exam formatters, TOS generators, the eOBE CLO manager. My failure pattern used to be identical to my student’s: I’d describe a form-filling task, Claude Code would immediately start guessing field names and dropdown logic, and I’d spend the next hour correcting assumptions instead of stating them upfront.
Running a grill-me-style pass first — forcing myself to answer “why does this field exist” and “what happens when the value is missing” before any code gets written — cut my SKILL.md iteration count roughly in half on the last two builds. The specification step matters more than the interrogation step: once the constraints are written down, I stop re-explaining them every session.
How to Install grill-me
npx skills@latest add mattpocock/skills
Select grill-me during setup. The repo’s install guide also recommends adding setup-matt-pocock-skills alongside it, since the full workflow (to-spec, to-tickets) lives in the same repository.
Takeaway: if Claude Code keeps handing you technically-correct solutions to the wrong problem, the fix isn’t a better prompt — it’s a rule that forbids the agent from starting until you’ve both agreed on the problem.
If you’re building your own Claude Code skills for teaching, client work, or internal tools, this four-rule brake is worth installing before your next session — it’s free, it’s one command, and it will save you the rework.
FAQ
What is the grill-me skill for Claude Code?
It’s a slash-command skill by Matt Pocock that forces Claude Code to ask clarifying questions — one at a time, each with a recommended answer — before starting any implementation work.
Why did grill-me get 7 million downloads?
Because it solves a universal agentic-coding failure: AI tools start planning and building the moment you give an instruction, often solving the wrong problem entirely. A four-rule brake that prevents this resonated widely.
What’s the difference between grill-me and to-spec?/grill-me clarifies the problem through structured Q&A. /to-spec turns the agreed answers into a written specification that describes constraints — not a copy of existing code — so future tasks don’t inherit outdated assumptions.
Why does to-tickets split tasks by user function instead of technical layer?
Splitting horizontally (database, then API, then frontend) means you can’t verify anything works until final integration. Splitting vertically (a user registers, a member logs in) gives each task a complete, testable loop, so wrong assumptions surface immediately instead of at the end.
How do I install the grill-me skill?
Run npx skills@latest add mattpocock/skills, then select grill-me during setup. Also select setup-matt-pocock-skills to get the full to-spec/to-tickets workflow from the same repo.
Does this workflow apply outside of coding, like writing SKILL.md files?
Yes — the same brake works for any Claude Code skill-authoring task. Answering “why does this exist” and “what happens on missing input” before writing the SKILL.md logic cuts revision cycles significantly, based on my own exam-formatter and eOBE skill builds.