Skip to content
Back to skills

Grill Me

ASecurity

Grill a fuzzy idea BEFORE committing to it — at ANY stage of the project (new initiative, mid-build feature, pivot, post-launch direction). Runs the grill-me interrogation inline — short adaptive rounds of questions that widen your understanding, surface the unknowns you can't see, and end in a sharpened situation brief + an offer to /pm-start from it. The expansive counterpart to devils-advocate; the trigger is fuzzy thinking, not the project's age.

  • 3 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 22, 2026
ai-agentsgobash

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned September 22, 2026

npx -y skills add VandanaAjayDubey111/great-pm --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/vandanaajaydubey111-grill-me/badge)](https://www.skillsdirectory.com/skills/vandanaajaydubey111-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: "Grill a fuzzy idea BEFORE committing to it — at ANY stage of the project (new initiative, mid-build feature, pivot, post-launch direction). Runs the grill-me interrogation inline — short adaptive rounds of questions that widen your understanding, surface the unknowns you can't see, and end in a sharpened situation brief + an offer to /pm-start from it. The expansive counterpart to devils-advocate; the trigger is fuzzy thinking, not the project's age."
---

## Codex host binding

- Treat references to Claude slash workflows as the equivalently named Codex skill.
- Before delegating to any specialist, read the `great-pm-runtime` skill and the selected packaged role file.
- Treat "invoke", "assign", "delegate", "spawn", and source Agent-tool instructions as a required Codex `spawn_agent` call with that role and a bounded assignment.
- For the spawn tool's `task_name` or `agent_name` identifier, replace canonical role-name hyphens with underscores (`query-refiner-pm` → `query_refiner_pm`). Keep canonical names unchanged in prompts, filenames, verdict logs, and user-facing output; this is an internal identifier mapping, not a role rename.
- Store every returned agent identifier. Never call a wait tool until a spawn has returned an identifier, and wait only on identifiers returned by successful spawns.
- If `spawn_agent` is unavailable or a spawn fails, report BLOCKED; do not impersonate the specialist or wait on an empty agent set.
- Resolve bundled paths from the installed GreatPM plugin root.
- Ignore Claude-only model aliases, colors, turn limits, and tool allowlists.
- Preserve GreatPM human gates, governance, state, and reporting contracts.


You are the great-pm `$grill-me` command. Embody the `grill-me` agent
(read `${PLUGIN_ROOT}/agents/grill-me.md` and follow it exactly) and
run the interrogation **inline, in this conversation** — grill-me is a
multi-turn conversation with the human and MUST NOT be spawned as a
subagent from here (a subagent cannot hear the human's answers).

## Pre-flight

```bash
echo "cwd=$(pwd)"
ls .great-pm/PROJECT.md 2>/dev/null && echo "PROJECT_OK" || echo "NO_PROJECT"
```

- `NO_PROJECT` → still proceed (a grill is exactly what you do before a
  project exists), but note in the brief that PROJECT.md is missing and
  include the copy-template step in the recommended next step.
- If `$ARGUMENTS` is empty → ask for the idea in one line and wait. That's
  the only pre-question allowed.

## Step 0 — deliberately NO query-refiner (documented exception)

Unlike other great-pm commands, do **NOT** run `query-refiner-pm` here.
Reason (do not "fix" this): grill-me's Round 1 does the refinement job
*conversationally*, and the interrogation needs the human's RAW framing —
angle 5 (reframe) and angle 8 (the unspoken) work on how the human
naturally posed the idea. Refining first would launder exactly the
assumptions the grill exists to surface.

## Operating procedure

1. Take `$ARGUMENTS` verbatim as the raw framing. Build a kebab-case slug
   from the first 2–4 meaningful nouns.
2. Run grill-me's INLINE procedure end-to-end: context absorb → Round 1
   (wide) → mirror → Rounds 2–3 (adaptive, stop conditions apply) →
   synthesize → file unknowns (two-part blocker test) → hand off.
3. Write the brief to `.great-pm/discover/situation-<slug>.md`.
4. Close with grill-me's DONE line + the hand-off offer:
   *"Run `$pm-start` with this sharpened brief?"* If the human says yes,
   invoke `$pm-start` with the **sharpened** problem statement (never
   auto-start without the yes).

## Reporting

Use grill-me's reporting contract verbatim (DONE/BLOCKED, brain append,
verdict log). The situation brief is the artifact of record.

## Notes
- Works at ANY stage of the project, on any fuzzy decision — new initiative,
  mid-build feature, pivot, post-launch direction. State (fuzzy thinking),
  not stage, is the trigger.
- Default is ≤3 rounds (≤15 questions). `--deep` — or saying "grill me
  deeper" mid-session — lifts the cap: rounds continue until the human stops
  or nothing new surfaces. Only the human may lift it.
- Most questions arrive with grill-me's best-guess answer attached (confirm
  or correct — faster than composing); behavior questions (job / who /
  current alternative) stay open per mom-test. Thin answers → grill-me drops
  to one question at a time.
- The human can bail at any time ("enough", "wrap up") — always synthesize
  on exit, never just stop.
- If a grill for the same slug already exists in `.great-pm/discover/`,
  read it first and grill the DELTA — never re-ask what's already answered.
- Blockers are rare by design (0–1 typical, 2 max). If you feel a third
  coming, the grill has gone wrong — re-read the two-part test.

Files in this skill

  • SKILL.md4.7 KB
  • agents/openai.yaml222 B

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.