By name: grill-me description: Ephemeral grill mode (no durable artefacts). PREFER /ro:grill-with-docs as the default for ANY grill since it captures CONTEXT.md + ADRs alongside the interrogation. Use /ro:grill-me only when (a) the user explicitly asks for a no-docs grill, or (b) the repo has nowhere sensible to write durable docs (e.g. a throwaway scratch directory). Both skills walk the same decision tree the same way; the only difference is whether decisions get written down. Triggers on "grill me ephemeral", "quick grill no docs". For everyday "grill me / let's build / before we code / pressure-test this / interview me about", grill-with-docs is the default. allowed-tools: Bash Read Glob Grep
Grill Me
PREFER /ro:grill-with-docs by default. This skill is the ephemeral version: same interrogation, no durable artefacts. The transcript dies with the session. Almost every grill is more useful WITH docs (CONTEXT.md + ADRs survive into the next agent's context window), so reach for grill-with-docs first. Only use this skill when the user explicitly says "no docs" or the repo can't write to a docs/ folder.
Thin facade around Matt Pocock's grill-me skill. The pattern: interview the user relentlessly about a plan or design until shared understanding is reached, walking the decision tree one branch at a time, resolving dependencies in order.
Bootstrap (idempotent)
Before grilling, ensure the upstream skill is installed globally so /grill-me works in any project, not just this one:
test -e ~/.claude/skills/grill-me || npx -y skills@latest add mattpocock/skills/skills/productivity/grill-me -g
Run this once at the top of the session. The npx call is a no-op if Matt's skill is already symlinked.
How to grill
- One question at a time. Never a numbered list. The user is more likely to skim and confirm than write a long reply.
- Always include your recommended answer with the question, and a short why. Skim-and-confirm is the unit of progress.
- Walk the tree. Resolve foundations before details. Choices that gate later choices come first.
- Explore over ask. If a question can be answered by reading the codebase or docs, do that instead of asking the user.
- Stop when shared understanding is reached, not when the user runs out of patience. If you sense impatience, summarise the resolved decisions and offer to switch out of grill mode.
When to switch out
The user says "enough", "let's start", "go AFK", or pivots to a different task. Hand off cleanly: list resolved decisions, flag what's still open, then proceed to implementation.
Meeting-transcript variant
Same pattern, different input. If the user pastes (or points at a file containing) a recorded meeting transcript, run the grilling over the transcript instead of starting from a fresh brief:
- Read the transcript end-to-end.
- Extract the decisions that were made, and the gaps where decisions should have been made but weren't.
- Grill the user on the gaps only. Skip questions the transcript already settled.
This is how post-meeting "we kind of agreed" becomes a written contract before any code is touched. Pairs with /matt-pocock-coding-workflow step 1.
Conceptual background
The pattern (interrogation precedes specification) is documented at llm-wiki-ai-research:grill-before-spec. The skill is one implementation of that pattern; the principle holds regardless of tooling.
Why this skill exists in ronan-skills
Two reasons:
- Bootstrap. Anyone who installs the
roplugin gets Matt Pocock's upstream skill for free on first invocation, without having to know about it. - Better triggers. The upstream
descriptiononly fires on the literal phrase "grill me". This one nudges Claude into interview mode whenever the user is about to build, design, or spec out something new.