grill

Read-only pre-design interview that surfaces tacit intent before planning. Use on explicit request, or before design when intent risk stays high after discovery and at least two hold: large blast radi

SKILL.md

GitHub

Interrogate the user to extract tacit intent before the real work begins. You are the questioner; the user is the validator.

Subject: $ARGUMENTS (if empty, ask what to grill about)

System — design's read-only pre-step

grill is an optional read-only front-step to design — not a pm-* loop node (it owns no per-task artifact, so the canonical loop is unchanged). It consumes the same per-task inputs design reads — pm-context links, retro memory, the pm-roadmap item — interrogates the user to resolve what they leave unstated, then hands the aligned intent to design. It persists nothing: the intent flows into the next step (design plan Decisions, or retro task memory). If no design/retro follows, the intent is intentionally session-volatile — by design, not a bug.

Inputs — read BEFORE asking (do not ask what these answer)

pm_root="$HOME/.config/ai/skills/pm-roadmap"
repo_root="$(git rev-parse --show-toplevel 2>/dev/null)" || repo_root=""
if [ -n "$repo_root" ] && [ -x "$pm_root/node_modules/.bin/tsx" ]; then
  pm() { PM_ROOT="$repo_root" "$pm_root/node_modules/.bin/tsx" "$pm_root/pm-roadmap.ts" "$@"; }
fi
  • Resolve the originating session first, using the exact injected metadata: PM_SESSION_TOOL="<tool>" PM_SESSION_ID="<exact id>" node "$HOME/.config/ai/lib/worktree.mjs" resolve-session --root "$repo_root" --tool "<tool>".
  • When it returns ok, run pm get <resolved plan id> and use its owning <KEY> to read .agents/tasks/<KEY>/links.md and memory.md, both read-only. An explicitly supplied item id may be passed directly to pm get.
  • Always grep/read the codebase for anything answerable from code.
  • An unbound session, no Task linked to its validated plan, or outside a repo → skip the PM probes and inspect code only. Never fall back to a launcher selection. Never run npm install from grill.

Protocol

  1. Explore first. Answer everything you can from the inputs. Only what code+context cannot answer goes to the user.
  2. One question at a time via AskUserQuestion. Never batch.
  3. Each question carries your recommended answer first + a one-line rationale.
  4. Dependency-tree order — each answer narrows the next; prune branches a prior answer closed.
  5. Hunt tacit knowledge — prioritize constraints / priorities / "why this way" over surface choices.
  6. Default 1 round; hard cap 3 rounds.

Stop conditions

Stop when ANY holds: the actionable choices are settled; the user says proceed/skip; questions start repeating; further questions no longer reduce design risk; the 3rd round is reached. Then emit a compact Extracted Intent bullet block for the next step to consume.

Hard rules

  • Write nothing. Write/Edit are omitted from allowed-tools, and Bash is restricted to read-only commands only: git rev-parse, git status/log/branch, pm list|tree|get|next|recent|validate|current-task, and read-only rg/grep/find/ls/sed -n. Forbidden: shell redirection, heredocs, tee, sed -i, npm install, and every pm write subcommand (task/add/plan/approve/close/drop/triage/memory/links/persist/complete/migrate --apply). Durable capture is delegated downstream — point there, never write it yourself.
  • Don't re-ask what code/context answers. Exploring first is mandatory, not optional.
  • Recursion SKIP: if already inside another skill flow (design / verify / …), do not start grill anew unless the user's latest message explicitly asks for it.
  • Questions are decision-shaped (about intent), not a code review.