Skip to content
Back to skills

Grill Me

ASecurity

Use when authoring an initiative, clarifying interdependent open questions, or designing a multi-quarter roadmap — depth-first one-at-a-time interrogation with agent recommendation + reasoning per question. Skip for bug-fix / implement / verify (too much friction).

  • 2 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 9, 2026
ai-agentsgit

Works with

  • cli

Security analysis

A100/100

Scanned October 9, 2026

npx -y skills add qyndex/claude-init --skill grill-me --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Grill Me?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Grill Me
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/qyndex-grill-me/badge)](https://www.skillsdirectory.com/skills/qyndex-grill-me)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: grill-me
description: Use when authoring an initiative, clarifying interdependent open questions, or designing a multi-quarter roadmap — depth-first one-at-a-time interrogation with agent recommendation + reasoning per question. Skip for bug-fix / implement / verify (too much friction).
when_to_use: When authoring an initiative or clarifying ≥3 interdependent open questions; depth-first one-at-a-time interrogation with agent recommendation per question.
model: sonnet
disable-model-invocation: false
---

# Grill me

> Vendored from [mattpocock/skills/skills/productivity/grill-me](https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md) (commit `bd04394c`). House-extended with recommendation format, codebase exploration policy, complexity gate, and AUTOPILOT bailout.

## When to invoke

Auto-on:
- `/initiative create` (every initiative, regardless of size)
- `/specify` for specs with `complexity: L` or `complexity: XL`
- `/clarify` when ≥3 `[OQ]` items cluster around one subsystem (interdependent)
- `/okrs` (when defining new KRs — they cascade)
- `/roadmap` (Now/Next/Later cuts — prioritisation cascades)

Hard-exclude (do NOT grill):
- Bug-fix / `/debug` / `/implement` / `/verify` — interrogation is too late
- Trivial specs (`complexity: S`) — batched questions are faster
- Refactors that don't change architecture
- Anything in `[autopilot]` / Cloud Routine / `--bg` context (no human to answer — use recommendations as defaults instead; log open questions in handoff)

## The core directive

Interview the operator relentlessly about every aspect of this plan until you reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one.

**For each question**:
1. Pose ONE question (never batch — one at a time)
2. Provide YOUR recommended answer
3. Give 3-4 multiple-choice alternatives
4. Explain WHY you recommend the option you did, in one sentence

**If a question can be answered by exploring the codebase, explore the codebase instead.** Never ask the operator "what does X do" when grep can tell you in 10 seconds.

## House format

Use `AskUserQuestion` with the `header` field. Format:

```
Q<N>: <one specific decision-forcing question>

  Recommended: <option you'd pick>
    Why: <one sentence reasoning — cite codebase evidence where possible>

  Other options:
    (a) <alt 1>
    (b) <alt 2>
    (c) <alt 3>
```

Pass the recommendation as the FIRST option in the `options` array (with "(Recommended)" in the label).

## Branch-walking discipline

Ask in dependency order. If Q2's answer depends on Q1, ask Q1 first. The mental model is a directed acyclic graph; you walk depth-first.

When you sense the operator is uncertain, ask a meta-question:
- "Should I keep grilling, or would you rather defer the rest to /clarify later?"
- "I have ~5 more questions in this branch. Continue, or stop here and let the rest emerge from prototyping?"

## AUTOPILOT bailout

When running in unattended context (`AUTOPILOT=1`, Cloud Routine, `claude --bg`, `claude -p`):

1. Do NOT call `AskUserQuestion`.
2. For each question that would be asked, write the recommended answer to a draft, with confidence: high|tentative.
3. Append the question + recommendation to `[OQ-pending-operator-review]` items in the spec/initiative.
4. Tag the artifact with `status: draft-autopilot-needs-review` (not `approved`).
5. Surface the count of unresolved Qs in the handoff (`OVERNIGHT_REPORT.md` or stream NEXUS handoff).

Never block on a human in autopilot. Never auto-approve as if the human answered.

## Closing the grill

When the design tree is exhausted (no more questions whose answers materially shape the artifact), stop. Don't ask filler questions to seem thorough.

Output a one-line confirmation:
> "Grilled. Drafting <spec/plan/initiative> now."

Then proceed.

## Hard rules

- One question per turn. Never batch.
- Always provide a recommendation. "What do you think?" is not grilling.
- Always provide reasoning. The operator should learn from the recommendation even when they override it.
- Cite the codebase when possible — grep is free, operator time is not.
- Don't grill beyond the depth that materially affects the artifact. Stop when stopping.
- Confidence: high for high-stakes branches, tentative for nice-to-have details. Make it explicit so the operator knows where to apply scrutiny.

## Attribution

Original: [mattpocock/skills](https://github.com/mattpocock/skills) — `skills/productivity/grill-me/SKILL.md` at commit `bd04394c675ee54173a093c50eb74da01a2940fa`. MIT-licensed.

House extensions (recommendation format, complexity gate, AUTOPILOT bailout, codebase-exploration substitution) are project-specific to this harness.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.