| name | grill-me-to-prd |
| description | Fully automated design session — two subagents interview each other to exhaustively explore a feature idea, then synthesize the dialogue into a PRD. Use this skill whenever the user describes a feature, system, or idea they want to build and says "grill-me-to-prd", "auto-grill", "run the pipeline on this", or "turn this into a PRD automatically". Also trigger when the user says they want to skip the manual grill-me session and go straight to a PRD from a raw idea. |
| compatibility | {"tools":["bash","read_file","write_file"]} |
grill-me-to-prd
Two subagents converse with each other — one interrogates, one answers — producing a rich dialogue and reasoning trail that feeds directly into a PRD. You are the orchestrator: you manage the session files, sequence the turns, and fire the synthesis at the end.
Directory layout
.claude/skills/grill-me-to-prd/
├── SKILL.md ← you are here (orchestrator)
├── agents/
│ ├── interrogator.md ← interrogator persona + instructions
│ └── answerer.md ← answerer persona + instructions
├── references/
│ └── prd-template.md ← PRD output schema for synthesis
└── session/ ← runtime scratch, gitignored
├── dialogue.md ← shared transcript (orchestrator-owned)
├── reasoning-interrogator.md ← interrogator's live reasoning state
└── reasoning-answerer.md ← answerer's live reasoning state
Read agents/interrogator.md, agents/answerer.md, and references/prd-template.md before starting.
Orchestrator process
1. Initialise the session
Create the session directory and seed all three session files:
session/dialogue.md — seed with the user's idea:
# Dialogue
## Idea
<user's idea verbatim>
---
session/reasoning-interrogator.md — seed empty:
# Interrogator reasoning
## Settled decisions
(none yet)
## Current focus
Starting — need to establish scope and constraints first.
## Open threads
(all — session just started)
session/reasoning-answerer.md — seed empty:
# Answerer reasoning
## Assumptions made
(none yet)
## Tradeoffs rejected
(none yet)
## Open questions I flagged
(none yet)
Tell the user: "Session started. Running the grill — this will take a few minutes." Then begin the loop immediately, no further input needed.
2. The loop (max 12 rounds)
Each round is two half-turns: interrogator then answerer.
Half-turn A — interrogator
Invoke the interrogator subagent with this exact prompt structure:
[SYSTEM]
<contents of agents/interrogator.md>
[USER]
<contents of session/dialogue.md>
---
<contents of session/reasoning-interrogator.md>
---
It is now your turn. Ask your next question, or output exactly "DONE" if the design is fully resolved.
The interrogator returns either a question string or DONE.
Half-turn B — answerer
Invoke the answerer subagent with this exact prompt structure:
[SYSTEM]
<contents of agents/answerer.md>
[USER]
<contents of session/dialogue.md>
---
<contents of session/reasoning-answerer.md>
---
It is now your turn. Answer the latest question in the dialogue.
Append to session/dialogue.md:
**A:** <answer>
---
Overwrite session/reasoning-answerer.md with the answerer's updated reasoning block.
Progress reporting
After each full round, print a one-line status to the user:
Round N/12 — Q: <first 60 chars of question>...
3. Synthesise the PRD
When the loop ends (DONE or 12 rounds), invoke a synthesis subagent:
[SYSTEM]
You are a senior product engineer. Your job is to produce a PRD from a completed design dialogue.
Do NOT ask any questions. Synthesise everything you know into the PRD template below.
<contents of references/prd-template.md>
[USER]
# Dialogue transcript
<contents of session/dialogue.md>
# Interrogator's reasoning trail
<contents of session/reasoning-interrogator.md>
# Answerer's reasoning trail
<contents of session/reasoning-answerer.md>
Produce the complete PRD now.
Write the output to session/prd.md.
4. Clean up and present
Print the full PRD to the user inline.
Offer two options:
- "Looks good — save it" → copy
session/prd.md to the repo root as PRD-<slug>.md
- "I want to adjust something" → accept their correction, patch the relevant section of
session/prd.md, reprint
Archive the session:
mv session/ sessions/$(date +%Y%m%d-%H%M%S)/
Rules
- You are the only writer to
session/dialogue.md. Agents only read it.
- Each agent overwrites (not appends) its own reasoning file every turn — the file always reflects current state, not a log.
- Never skip a round because an answer seemed complete. Let the interrogator decide when it's done.
- If an agent invocation fails or returns malformed output, retry once, then skip that half-turn and note it in
dialogue.md as [SKIPPED — agent error].
- The session directory is ephemeral. If the user runs the skill twice, initialise a fresh session each time.
