human-reviewlisted
Install: claude install-skill victorrentea/human-review
# /human-review — assemble a review.html for a human
**This skill does not review the code. It writes up a review that already happened.**
The review happened in the coding agent's own conversation, and it left a record:
`/record-review`, run in the conversation that implemented the change, reviews it
adversarially (`/code-review` in Claude Code, reviewer subagents elsewhere) and writes
**`review-points.md`** at the repository root — what it fixed because the review was right, what it read and **declined**, and what it **assumed** where the ticket was
ambiguous — committed with the fixes, so the record arrives in the pull request's own file
list. This skill reads that file. It does not run a review, and it does not write one.
That is the whole division. The judgement is produced once, by the agent that made the
decisions, while it still has them; the page is assembled by programs from the branch.
What is left for you is two model-written artifacts — the cheap model's check of the
sentence↔test pairing behind the Tests tab (`test-mapping.json`: every scripted link
confirmed or rejected, the rest paired), and **+1 LLM script**, the Demo film's — plus the page's layout and its ledes. Everything else on
these five steps is a script. Do **not** commit or push.
A branch with no `review-points.md` is not an error and not a gate: the page says so, in a
band, and the piles read *not recorded*. **Do not write the piles by hand to fill the
gap** — a pile you compose at the end of a