aidoc-flow-framework: Skill for Claude Code

Pre-production review

Muse project-skill instance of the canon preprod-review skill. Source: aidoc-flow-claude-agents-config/skills/preprod-review/SKILL.md. Engine-neutral wording mirrors framework/skills/preprod-review/. Skill path in this repo: .agents/skills/preprod-review/. Where the canon names a Claude-specific agent type, use this harness's focused read-only subagent equivalent and keep the independence contract.

Reviewing something well before it ships is not one read — it is several independent reads from angles that a single reviewer conflates and therefore misses. A security lens and a docs lens hunting the same diff surface different truths; fused into one pass they blur. This skill fixes the lenses, runs them in parallel against real source, and forces every surviving finding through a verification step so the verdict is evidence, not vibes.

Use it before anything that widens blast radius:

• cutting a release or version tag
• promoting a library, template or workflow to a team/company default
• rolling a change out to N consumer repos
• a deploy or infrastructure change that is hard to reverse
• handing an artifact to adopters who cannot see its internals

Not for ordinary in-flight review of a PR or a work-in-progress branch — that is /code-review. This is for the ship/elevate/rollout decision, where the cost of being wrong is paid by people who were not in the room.

The shape

1. Scope the surface (you, inline — a few minutes). List every file the review must cover, then read the 2–3 most privileged or most load-bearing ones yourself, so you can spot-check findings later. What counts varies by artifact, but the shape is always entry points, whatever they trust, and the surfaces adopters read:

   Artifact	Typically
   CI / workflows	.github/workflows/*.yml, install or distribution templates, the scripts they invoke, adopter docs
   library / service	public API surface, config and defaults, migration steps, the release or deploy script
   infrastructure	the changed resources, IAM/permission boundaries, the state and rollback path
   any of them	the governance surfaces — CHANGELOG / ROADMAP / decisions log / handoff

2. Fan out five lenses in parallel (one dispatch, five agents — see below). Each gets a scoped brief and returns a ranked report. Running them in one turn is the point: they finish together and don't inherit each other's framing.

3. Synthesize + verify (you, inline). Dedup across the five reports. For every BLOCKER and every HIGH, open the cited file:line and confirm the defect is real before it survives into your output. A plausible-sounding finding that doesn't reproduce against source is noise — cut it or mark it PLAUSIBLE with what would confirm it.

4. Deliver one ranked verdict: BLOCKER (do not ship), SHIP-WITH-FIXES (fixable, enumerated), or READY. Lead with the verdict, then the ranked findings, then strong positives to not regress.

5. Promote surviving findings (you, inline — the orchestrator owns this). Lenses must not file; their reports are raw synthesis input. Check every surviving BLOCKER and HIGH against the submit-feedback promotion bar and file the ones that clear it via submit-feedback (search-first, body contract, safe publication). Link the issue numbers from the verdict so no finding dies at the transcript boundary.