Shape Idea

Keep alignment separate from delivery

Shaping settles decisions; implementation applies them. Do not change product source, configuration, or dependencies, even for an edit you plan to revert. A changed line mixes alignment with delivery and leaves unreviewed code behind.

Write only to the spec folder, glossary, decision contracts, and vendor agent context. Run every experiment, benchmark, and preview in a scratch directory outside the working tree. If no experiment can answer a question, record an assumption and its risk. If code contradicts the user or a decision, surface the conflict; do not fix the code.

Check the working tree when you start and again before writing spec.md. Revert what this session changed outside the allowed paths and keep only what you learned. Leave uncommitted work that predates the session alone.

Ground decisions in project truth

Before the first question, invoke project-knowledge and apply it throughout the session. If it is unavailable, read GLOSSARY.md and relevant subjects from docs/decisions/README.md when present, update confirmed terms, and surface terminology or decision conflicts for explicit clarification.

Read root PRODUCT.md when it exists before settling the work unit. Treat it as the current app-level premise, use only the product constraints relevant to the selected work, and surface any mismatch that would require changing that premise. Do not create, edit, or copy the whole file into the work-unit spec. Missing PRODUCT.md does not block shaping.

Resolve what available evidence can answer before asking the user.

Ground any conclusion about a third-party package or tool in evidence of how it actually behaves — its own source, documentation, releases, and maintainer statements — and confirm it against this project's versions in a scratch copy before building on it or working around it. Record what was checked, what fell short, and the upstream change that would reopen the decision.

When a decision selects a framework or hosted service, invoke add-stack-context when available and let it own discovery, source acceptance, installation, live vendor-document routing, and accounting. When it is absent, retain the selected technology's outcome inline: use find-skills when available, or an equivalent current Skills search when not, verify vendor control before installing an official skill, require explicit approval for a community skill, and check other official vendor channels. Keep a changing llms.txt at its official source and preserve a bounded AGENTS.md or CLAUDE.md instruction that retrieves it during relevant work instead of copying its contents into the repository. Account for the selected technology before continuing.

Present one decision at a time

Present a concrete candidate for the user to correct, using the lightest medium that makes the decision judgeable.

• Decide an inexpensive, reversible choice when a mismatch is unlikely or easy to detect; state it as an overridable assumption, never a project decision contract.
• For a branch expensive to get wrong, ask exactly one question about one fact, value, or choice. Include a recommended answer and concise reason, then wait.
• If a proposed decision depends on information only the user can know, state that information and ask whether it applies. Verify any condition you can check yourself.
• For a choice judged by looking or trying, inspect the current surface as evidence and show it only when the decision requires a baseline comparison. Render a candidate or two or three controlled variants, verify the relevant states, and wait for the user's reaction. Invoke build-prototype for a whole-surface review or a temporary comparison for the user's specific situation. Scope that decision layer independently of product screen count: pass only the question and context needed to judge it, integrate the chosen result into the existing prototype when present, discard the comparison, and resume shaping. If no sufficient renderer is available, defer the decision and record the resulting risk.
• When a flow, state model, or relationship has multiple branches, transitions, or links, render one diagram before a downstream decision. Ask at most one question about its unresolved part and wait. Keep a linear structure that fits in one sentence in prose.
• When the user asks for an explanation rather than a decision, invoke explain-visually. If unavailable, use one sentence when sufficient or the best available renderer otherwise.

A choice is settled when the user confirms it or it is made under authority the user explicitly delegated for that class of decision. It becomes a project decision contract only when future work should reuse it, its rationale prevents reasonable re-litigation, and it came from a real trade-off; feature-local choices stay in the spec.

Skip review only for an already confirmed pattern, routine presentation details, or explicit user delegation. Record the reason and treat only agent-judged reasons as assumptions.

Write the product contract

Stop asking questions when every implementation-relevant decision is resolved or explicitly deferred; do not wait for the user to declare completion. Translate confirmed product-change requests into required behavior. Keep cheap agent-chosen defaults as overridable assumptions; ask about or explicitly defer consequential unsettled behavior and record its possible impact as a remaining risk. Never record a deferred branch as a settled constraint. Interim behavior the product needs while it stays open is written under the deferred point, naming the decision that replaces it; where the contract states that behavior again so it can be built and tested, mark it interim there too.

When ready for implementation, write docs/specs/<slug>/spec.md as the stable product contract, creating the kebab-case folder when needed. Include the user-visible outcomes, approved scope, observable acceptance criteria, settled constraints and rationale, assumptions, off-limits areas and why, deferred points, and remaining risks. Record behavior and decisions without predicting files, functions, code structure, technical layers, or implementation steps. Carry only applicable app-level constraints from PRODUCT.md; keep the file as their canonical product context rather than duplicating its full contents. Summarize the same contract and do not prompt for another action.