critic-looplisted
Install: claude install-skill waseemnasir2k26/skynetlabs-all-claude-code
# critic-loop
LLM-as-judge quality gate. Operationalizes the "2nd agent grades/self-corrects
before you accept" lesson (claude-code-mastery, video 8). SkynetLabs only had the
manual `/review`; this is the automated gate that runs before a deliverable ships.
## Core principle
**The builder never grades its own work.** A model that wrote an artifact is biased
toward approving it. Spawn a SEPARATE critic agent with a clean context, give it the
artifact + rubric + adversarial framing, and let IT decide. Loop until PASS.
```
build → spawn critic (FAIL-biased) → PASS? ship : apply fixes → re-critique
↑________________________________________|
max N rounds (default 3)
```
## When to use
USE before:
- A client deliverable is sent (website, copy, audit, proposal, automation)
- A deploy / publish / `git push` to production
- Anything with secrets, credentials, or money flow
SKIP when:
- Throwaway scratch work, internal notes, exploratory spikes
- Already passed critic this session AND artifact unchanged
- Time-critical hotfix where the fix is trivial + reversible (still log it)
## Inputs required
1. **Artifact** — path(s), URL, pasted text, or git diff (`git diff main`)
2. **Type** — one of: `website` | `copy` | `automation` | `proposal` | `generic`
3. **Rubric** — auto-selected from type (below), or user-supplied for `generic`
4. **Mode** — `single` (1 critic, default) | `multi` (3 lenses, majority vote)
5. **Max ro