Think: Reason Before You Build

Outcome Contract

• Outcome: a rough question becomes one defensible recommendation whose weakest assumption is named.
• Done when: the goal, constraints, chosen direction, the rejected alternative and why, the most fragile assumption, and the next step are stated; nothing load-bearing is silently assumed.
• Evidence: repo state, docs and ADRs under docs/decisions/, live config values, primary-source docs, and explicit user preferences. Memory is a lead to re-verify, never evidence.
• Writes: none — no code, plan, or report. Keep a result via brainstorm lock; an approved build goes to to-plan or work.

Modes

Resolve from a leading subcommand, else the request.

An error, crash, failing test, or "why is this broken" → one line: it belongs to hunt; route there. Fuzzy intent → grill.

Process (decide)

1. Frame. The decision in one sentence and what "good" means (success signal, constraints, who pays the cost). Sources conflict, or two readings cost differently → name the conflict and ask which wins; never silently pick.
2. Ground before opining. Read AGENTS.md/CLAUDE.md and only the matching rule or ADR. Open the real config for any default, env var, or setting the answer depends on; never quote one from memory. Separate facts from decisions: find facts yourself (repo, docs, sub-agents for wide searches); put only decisions to the user.
3. Official and proven first. Check built-ins and ecosystem standards in live primary docs; the official solution wins unless you can say why it falls short. For a hard or repeatedly-tuned problem, study 2–3 mature projects and name what you take from each.
4. Generate real alternatives. Genuinely different, not variations, beside the brute-force option (simplicity-gate). Interface shape is the crux → switch to twice.
5. Run the lenses. Pick 3–5 that fit from references/lenses.md; its Core four almost always apply (e.g. attack angles for external dependencies, scale, data migration; deletion test, depth/seam for module design; entity delta when the plan adds settings, flags, commands, or services). One sentence per lens finding; nothing found → one line.
6. Deform or discard. A lens finds a hole → change the design to survive it; it shatters the approach → drop it and say why. Never hide a failed lens.
7. Recommend. One direction: effort, risk, and the existing code it builds on. Name one alternative only if the call is close (>40% the user would prefer it).

A question only running something can settle (does this state model feel right?) → recommend a throwaway prototype answering exactly that; building it needs the user's go-ahead. A reversible call inside the approved scope is decided and reported, not asked.

Grill-lite

Decisions only the user can make → references/grill-lite.md.

Hard Rules

• No placeholders. "To be decided", "details later", "similar to step N" mean the thinking is unfinished.
• Zero-setup default. Anything every user must install or configure (hook, MCP server, config key, new runtime or language) first says why a built-in or fixed default cannot do the job. A second language or runtime in a single-stack project needs explicit approval.
• Phases stand alone. After phase N ships, the system works even if N+1 never lands; "Phase 0: investigate" belongs in this thinking, not the plan.

Gotchas

• Design rejected → ask what specifically failed; re-enter with narrowed constraints, never from scratch.
• A rework that would remove the product's differentiator → check docs for a deliberate choice; deliberate → Keep.

Output

Verdict:        [one recommended direction, one sentence]
Why:            [2–3 reasons tied to this user's constraints, not generic trade-offs]
Rejected:       [the closest alternative and why it lost]  (omit if none was close)
Assumes:        [most fragile assumption → what happens if it fails]
Lenses:         [lens → finding, one line each]
Open:           [decisions only the user can make, as a frontier]  or  [none]
Next:           [the one concrete step: brainstorm lock / to-plan / work bounded / prototype / hunt]

evaluate and triage use their reference's format instead. Keep prose short; the block is the answer.

Defer to: grill (fuzzy intent), brainstorm (lock), to-plan (locked plan), hunt (errors, regressions), check (built code).