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.mdai/skills/grill/SKILL.md
GitHubInterrogate 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, runpm get <resolved plan id>and use its owning<KEY>to read.agents/tasks/<KEY>/links.mdandmemory.md, both read-only. An explicitly supplied item id may be passed directly topm 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 installfrom grill.
Protocol
- Explore first. Answer everything you can from the inputs. Only what code+context cannot answer goes to the user.
- One question at a time via
AskUserQuestion. Never batch. - Each question carries your recommended answer first + a one-line rationale.
- Dependency-tree order — each answer narrows the next; prune branches a prior answer closed.
- Hunt tacit knowledge — prioritize constraints / priorities / "why this way" over surface choices.
- 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-onlyrg/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.