visual-recap
Visual Recap
Render a finished change as a single self-contained interactive HTML file that shows a reviewer the shape of the diff before they read the lines: which surfaces moved, which hunks are load-bearing, what's risky — with callouts anchored to the exact lines, then a drop into the literal code. The output is reviewer comprehension as a portable artifact, not a defect list.
It renders in one of two modes, chosen in Step 1 and never defaulted into: a scrolling recap when the reviewer is auditing a change, and a walkthrough deck — one idea per screen, premise → changes → mechanism → aftermath — when the reviewer has to be argued through one.
This skill exists because Cohen's strongest empirical finding is author preparation: an annotated walkthrough prepared by someone who understands the change correlates with near-zero defect variance. The pipeline produces that annotation as terminal prose (/pre-merge, /walk-commits) but never as a highlightable, shape-first surface — so the reviewer rebuilds the change's topology in their own head before judging any line. /visual-recap produces the missing surface (Meadows: a missing high-bandwidth feedback flow in the review loop) without any lock-in — the agent authors the HTML directly; there is no renderer package, no build system, no server, no MCP connector, and no hosted database. GitHub PR review comments stay the durable channel.
It also renders the forward direction — a plan's decisions and architecture, before any of it is built. There is no separate /visual-plan skill and none is coming (#317): a second command name buys nothing the mode selection in Step 1 does not already buy, and costs an inventory surface in every discovery doc. The two directions share one data model — a recap renders from a finished diff, a plan renders toward a change that does not exist — which is why they share a token core, a serializer, and a grounding rule rather than a codebase.
Invocation Position
This is a side-route skill. It does not block merge and is never auto-invoked.
It reconnects to the main pipeline at the /pre-merge → merge boundary, exactly like /walk-commits: run it when you want shape-first comprehension before approving, then proceed to merge (and /compound if a durable lesson emerged). /pre-merge Phase 4 may recommend it — without invoking it — when the person merging didn't author the diff, the same way it recommends /walk-commits.
Use /visual-recap when: