Practical grill‑me Guide: Let an AI Agent Fully Clarify Requirements Before Coding

The article introduces the /grill-me user entry point and the /grilling interview primitive from the mattpocock/skills repository, explaining why they are split: /grill-me only triggers a session, while /grilling contains twelve hard‑coded rules that enforce a disciplined interview process.

What You Will See

The layered relationship between /grill-me (one‑line wrapper) and /grilling (12‑line rules).

A full simulated scenario: adding an export feature to an internal dashboard.

The five hard rules and the failure modes they prevent.

A decision‑matrix that tells you when to use /grill-me versus alternatives such as /grill-with-docs, /to-spec, /wayfinder, or domain-model.

Minimal prompt template and three launch methods (skills.sh installer, Claude Code plugin, other harnesses).

Matt Pocock’s recent update that discourages using /grill-me as a coding entry point.

Skill Hierarchy

The repository defines two files: skills/productivity/grill-me/SKILL.md – a 7‑line wrapper that can only be invoked manually ( disable-model-invocation: true, allow_implicit_invocation: false). skills/productivity/grilling/SKILL.md – a 12‑line model‑invoked primitive that enforces five hard rules: (1) ask one question at a time, (2) always provide a recommended answer, (3) separate fact‑finding (Agent) from decision‑making (User), (4) require shared‑understanding confirmation before any action, and (5) remain stateless.

These constraints prevent the agent from acting on unaligned requirements, a common source of rework.

Step‑by‑Step Scenario: Adding Export Functionality

Agent discovers facts : it inspects package.json (tech stack), looks for existing /export routes, reads db/schema.prisma for the data model, and checks documentation for any export‑related ADRs.

Question 1 – format : the agent suggests CSV (easy, no extra dependency) and offers XLSX as an optional upgrade, asking the user to accept CSV only.

Question 2 – fields : it recommends exporting all columns visible in the current board view, explaining why this avoids ambiguous field‑mapping.

Question 3 – filtering : it asks whether to respect the board’s current filter (e.g., only Beijing region) or to export the full dataset, offering a switch to ignore the filter.

Question 4 – filename : it proposes a naming convention {board‑name}-{YYYYMMDD‑HHmm}.csv and an alternative UUID‑based name for privacy concerns.

Question 5 – route reuse : after finding an existing placeholder route /api/reports/:id/export, it asks whether to reuse it or create a new /api/export endpoint.

Shared‑understanding gate : after eight to ten questions, the agent summarizes the decisions (format, fields, filter, filename, route) and asks the user to confirm before proceeding.

If the user confirms, the session ends; the agent does not write code until the gate is passed.

Extracted Hard Rules

Never ask multiple questions at once – avoid dependency interference.

Every question must include a recommended answer – keep the user from filling blanks.

Agent handles fact‑finding; the user makes decisions – clear separation of responsibilities.

Shared‑understanding confirmation is mandatory before any implementation.

The skill is stateless – no persistent files are created; only the conversation holds the shared understanding.

When to Use /grill-me

A decision table (described in text) classifies scenarios:

Temporary pressure‑tests or lightweight alignment – use /grill-me.

Projects needing documented decisions – switch to /grill-with-docs or domain-model.

Large work that exceeds a single session – use /wayfinder to build a decision map.

Simple bug fixes where decisions are already aligned – skip grilling and go straight to triage → TDD.

High‑fidelity UI/UX questions – do not use grilling; prototype first.

Launch Methods and Prompt Template

Three ways to invoke the skill:

A. skills.sh installer :

npx skills@latest add mattpocock/skills
npx skills update grill-me

then run /setup-matt-pocock-skills to configure the repo.

B. Claude Code plugin : add the plugin via the marketplace and install mattpocock-skills@mattpocock.

C. Other harnesses : any harness that respects the skills.sh installer can load the skill (e.g., Codex).

Minimal prompt template (to be placed in the wrapper):

/grill-me {your clear‑direction but undefined boundaries}

Constraints (recommended):
- You may act only after shared‑understanding confirmation.
- Facts are discovered by the agent; do not ask it to guess.
- Every question must include a recommended answer.
- Ask only one question at a time.

Recent Updates and Recommendations

Matt Pocock publicly announced that /grill-me is no longer recommended for coding entry; instead he suggests /grill-with-docs or the newer domain-model workflow for aligning plans with the codebase before implementation. The skill remains useful for narrow pressure‑tests, temporary alignment, or pre‑grill warm‑ups.

When Not to Use /grill-me

High‑fidelity design questions that require prototypes.

Workloads too large for a single session – switch to /wayfinder.

Pure implementation tasks where decisions are already settled – go directly to triage → TDD.

Situations needing persistent decision records – use /grill-with-docs instead.

When the harness already enforces the same hard rules – disable the grill trigger.

Question Volume, Dependency, and Gate Relationship

The three constraints of /grilling actively suppress ineffective questioning: they enforce decision‑tree progression, require recommended answers, and mandate a final shared‑understanding gate. A sudden explosion of questions (e.g., >540) signals that the session has lost control and should be paused or split.

Key Source References

All factual anchors come from the mattpocock/skills repository (e.g., SKILL.md files, .agents/invocation.md, changelog entries) and from Matt’s update page on AI Hero. Community data such as typical session length (16‑50 questions) and the upper bound of ~540 questions are quoted from Matt’s own articles and a Medium post.

Conclusion

The /grill-me wrapper launches the disciplined /grilling primitive, which ensures that facts are gathered, decisions are made by the user, questions are asked one at a time with recommendations, and a shared‑understanding gate protects against premature implementation. While still valuable for lightweight pressure‑testing, Matt now recommends using /grill-with-docs or domain-model for coding‑related workflows, reserving /grill-me for non‑coding alignment and quick validation.