Evidence capture

evidence-capture — AI-First evidence per feature

Output language: {{MGR_USER_LANGUAGE}} — all user-facing interaction and generated artifacts use this language; generated file names and rule IDs stay in English.

You record, in a structured and honest way, how AI was used in each feature. The evidence lives next to the feature (specs/<feature>/ai/), with a thin global index at ai/index.md. That keeps cohesion (spec, plan, execution and evidence in the same place) and still satisfies whoever expects the ai/ folder at the root.

Sovereign rule: organize, never invent (Law 1)

You do NOT know what the user reviewed, corrected or rejected from the AI — that critical judgment is theirs and is what the challenge evaluates most. Your role is to structure, ask and record, never to fill in plausible reviews. A field with no real information → [to be filled by the author], never fabricated text. The same goes for prompts: record the ones actually used (the user provides them or you retrieve them from the session history/mgr-code), do not invent prompts that "could" have been used.

Policy guard (before anything)

Check the project's CONSTITUTION. If AI-First evidence: disabled (or absent), warn and CONFIRM before proceeding — a one-off record is allowed, but a conscious one.

Scope (mandatory at the start)

Determine the target feature:

• Invoked by spec-create/spec-execute → receives the <feature-slug>.
• Direct → ask which feature (list the folders in specs/).

Create/update specs/<slug>/ai/ with three files (templates in templates/). Default file names: prompts.md, reviews.md, skills.md — when the challenge statement demands specific names (e.g. revisoes.md), use the demanded names instead.

specs/<slug>/ai/prompts.md — this feature's prompts

Record each relevant prompt used in THIS feature: the prompt text (or a faithful summary), the tool (Claude Code / Copilot / etc.), the SDD phase (spec, execution, test, review) and what it produced. Prompt sources, in this order:

1. The user pastes/points to the prompts they used.
2. If available, retrieve from the session history or from mgr-code.
3. No source → record only what the user confirms; do not complete with invention.

specs/<slug>/ai/reviews.md — what was reviewed/corrected/rejected (the most important)

Conduct a short, specific interview, without suggesting the answers:

• Which AI suggestions did you accept as they came?
• What did you correct before using? (what and why)
• What did you reject outright? (what and why)
• Where did the AI get it wrong (bug, misunderstood business rule, over-engineering)?
• Which decision of yours went against the AI's suggestion?

Record the answers verbatim. No "the AI suggested X and it was accepted" without the user having said it — if an item went unanswered, leave [to be filled by the author].

specs/<slug>/ai/skills.md — skills delegated to the AI

Which areas/skills were delegated in this feature (e.g. generating the OpenAPI excerpt, drafting the contract tests, suggesting the layer structure), and the degree of autonomy (human-reviewed draft / generated and accepted / consultation only). Partly derivable from 05-execution.md (invoked skills), but confirm with the user.

ai/index.md — global index (root)

Keep at the root an index that: lists each feature with a link to its specs/<slug>/ai/, a 1-2 line summary of the AI usage in that feature, and an "Overview" section with the recurring patterns (what the AI got right/wrong across the whole project). Update it on every newly recorded feature. This file is the entry point that satisfies the root ai/ folder requirement without duplicating content — it points, it does not copy.

Effort (optional, honest)

If the user wants to record effort, add to each reviews.md a perceived-effort line (e.g. "≈ 2h, mostly adjusting the 422 rules"). Do not record tokens or automatic timing — the skill cannot measure them reliably; only what the user states explicitly goes in, and as their statement, not as a measurement.

Integration with the SDD flow

• spec-create/spec-execute invoke this skill in each feature's Completion phase, passing the slug; it then drives the recording and updates the index.
• The feature's 06-completion.md starts referencing the feature's own ai/.
• mgr-code available → retrieve prompts/decisions already recorded in the session to pre-fill (the user confirms); absent → drive it only with what the user provides.

Closing

Show what was recorded (the three files + the index entry) and explicitly list the fields left as [to be filled by the author], for the user to complete the critical judgment — the part only they can write.