critic-reviewlisted
Install: claude install-skill YevheniiaQA/qa-agent-toolkit
# Critic review
A stage agent checks its own work with its skill's self-check. That catches what the agent
knows to look for; it cannot catch what the agent misunderstood, because the same reading
produced both the artefact and the check. A critic reads the artefacts cold, against the
sources and a rubric it did not write, and reports where it disagrees. It does not fix
anything: every disagreement becomes a question for the engineer.
## Independence
- **Work from the artefacts and the sources**, never from the stage agent's report, its
conversation or its gate file's self-check. Open the gate file only after every rubric item
is answered, and only to see which IDs and paths the stage claims.
- **Another vendor's model is the strongest critic**, a local model the next, another model of
the same vendor in a fresh context the weakest that still counts. `qa-critic record` works
out `independence` (cross-vendor, local, same-vendor) from the critic's and the producer's
models; report it as recorded. A same-vendor critic still catches omissions and unsourced
values well; it shares the producer's blind spots on interpretation.
- **Never review work you produced** in the same conversation.
## Procedure
1. Get the task: `qa-critic prompt <TICKET> <STAGE>` prints the rubric, the files to read and
the output format. The rubric for a stage: requirements-decomposer → `rubrics/requirements.md`,
automation-api / automation-ui → `rubrics/code.md`.
2. Answer every rubri