pi-grill-deck

An interactive question deck extension for pi, the AI coding assistant. It replaces the "dump a wall of markdown questions and hope the free-text reply maps back" pattern with a structured TUI: the agent hands all questions of a round to one tool, you answer them in a single screen, and the agent gets clean structured answers back.

Built for grilling-style sessions — design-tree interviews where each round presents the whole frontier of askable questions — but it works for any batch of questions (checklists, requirement gathering, triage).

Install

pi install npm:@evreke/pi-grill-deck

or project-scoped (shared with your team via .pi/settings.json):

pi install -l npm:@evreke/pi-grill-deck

New sessions pick it up automatically; /reload hot-loads it into the current one.

The experience

You ask pi to grill you about a decision. Instead of markdown, your editor is replaced by the deck — every question of the round on one screen, the focused one expanded with the agent's recommendation:

 grill deck · 5 questions · 1 answered · model-authored — verify before accepting
─────────────────────────────────────
  ✓ Q1 Scope: what does 'PR review format' actually cover?
        (a) + (b) + (c) — description template…
> · Q2 What is the primary pain driving this?
      What hurts today that this format must fix?
      ➡️ rec: (a) + (d) — vague descriptions and missing…
  · Q3 Scope of adoption: one repo or org-wide?

Answer with single keystrokes, submit, and the agent recomputes the frontier and opens the next deck. Deferred questions stay visibly open in the design tree and come back in a later round.

Keys

Key	Action
↑ ↓	move between questions
Enter	open the options view — recommendation pre-highlighted as option 1
a	accept the recommendation (quick path, no options view)
A	accept all recommendations at once
1–9	pick a numbered option (in options view)
e	write a custom answer (free text, saved verbatim)
s	defer — keep the question open for a later round
ctrl+s	submit manually — a fresh deck also auto-submits once every question is answered or deferred (a /grill revision session always uses ctrl+s so several answers can be changed first)
Esc	cancel the deck

/grill reopens the last deck with answers prefilled for review/revision.

For agents

Call grill_deck instead of printing questions as markdown: one call per round, containing the entire frontier — every question whose prerequisites are already settled. Each question: stable id, one-line title, optional body, optional choices, and your recommendation. The result marks every answer as accepted / choice / custom / DEFERRED (still open — re-ask it later with the same id). When the frontier is empty, summarize the shared understanding and wait for the user's confirmation before acting.

Skills included

This package bundles two skills derived from mattpocock/skills (MIT — see ATTRIBUTION.md):

• grilling-fork — the interview methodology: map decisions as a design tree, work the frontier in rounds, dispatch sub-agents for facts, don't act until shared understanding is confirmed. Adapted from the upstream grilling skill and renamed with the -fork suffix (see below).
• grill-me — the user-facing entry point ("run a grilling session").

Together with the bundled grill_deck tool this gives you the complete workflow out of one install: the skill drives what to ask and in which order, the tool presents each round as an interactive deck. The tool's prompt guidelines tell the model to use the deck for grilling rounds — the skill text ships with the package and is co-versioned with it.

Backward compatibility: the fork keeps the upstream methodology core verbatim — the design tree, working in rounds, the frontier definition, sub-agent fact-finding, the confirmation gate — and tries to stay drop-in compatible for a session: the same trigger phrases, the same session arc (rounds until shared understanding, then confirmation). It is nevertheless deliberately modified: rounds are presented through the deck, the round contract is stricter, and termination has an exploration guard. The exact inventory lives in the CHANGELOG under Skill divergence from upstream.

Working alongside the upstream grilling skill: the methodology skill is deliberately named grilling-fork, not grilling — if you also have Matt Pocock's grilling installed (e.g. straight from mattpocock/skills), the two coexist as separate skills and nothing needs to be removed. The pairing between the tool and the fork skill is enforced at three levels:

1. System-prompt guideline. While the package is installed, the grill_deck tool's guideline names grilling-fork as the session protocol and prefers it over other grilling-style skills — the binding holds no matter which skill a phrase-match would otherwise pick.
2. Description precedence marker. The fork's description says it is a deck-first fork and to prefer it over the plain grilling skill — so the choice is anchored already at skill-discovery level.
3. Explicit command. /skill:grilling-fork forces the fork deterministically (pi-native, always works).

The bundled grill-me keeps its upstream name; if you also installed it globally, your copy wins (pi keeps the first skill on a name collision and notes it at startup). You can toggle package resources explicitly with pi config.

Trust & security

Everything inside a deck — question text, choices, recommendations — is authored by the model. It is rendered as-is except for terminal escape sequences, which are stripped so they can't spoof UI state or touch your clipboard, and the deck header reminds you to verify before accepting. Rounds are capped at 32 questions / 2,000-character fields, and data replayed from the session file is validated before it reaches the UI or the model. The full security audit is maintained in the development repository.

How it works

Each submitted round is appended to the pi session as a custom entry (grill-deck-round, revisions as grill-deck-revision) and rebuilt on session_start — state survives /compact, /resume, and restarts.

Development

Pure logic (sanitization, session-replay validation, answer formatting) lives in lib.ts, unit-tested with npm test (Vitest). REGRESSION.md documents the interactive TUI checklist to run before a release.

Docs

Full user + agent onboarding: onboarding.md.

Compatibility

pi (interactive TUI mode). In non-interactive modes the tool degrades gracefully and instructs the model to print questions as markdown instead.