rubric-panellisted
Install: claude install-skill hsb3/dotfiles-agents
# Rubric panel
<!-- harness:claude-code -->
In Codex, first read [Codex distribution](../delegation/references/dispatch-knobs.md#codex-distribution)
for installed role setup, native dispatch, isolation prerequisites, and completion routing.
That section supplies the Codex procedures wherever this workflow names Claude Code tools.
<!-- /harness -->
Judge the artifact, not the process. Score absolutely against anchors,
never on a curve.
## Inputs (ask only for what's missing)
- **targets** — artifact path(s), implementation + tests per unit.
- **rubric** — anchored 1–5 dimensions with weights. None → offer a
default (cold readability 20; idiomatic fit, signal-to-noise,
architecture 15 each; type leverage, error-path robustness, test
quality 10 each; extensibility 5).
- **contract** — the behavioral spec; defines "externally observable"
for finding classification.
- Optional: panel size (default 3), personas (default: long-term
maintainer, language purist, QA skeptic), effort level.
## Preconditions
Run each target's gate first — never score broken code; report the gate
failure instead. Judges are read-only.
## Process
1. Dispatch N parallel judge agents, one persona each — one dispatch per
judge, all in a single turn so they run in parallel. Each is a
read-only reviewer subagent with its persona carried in the prompt.
**Every judge scores the whole field** — one judge across all targets
keeps calibration consistent; N judges average out persona bia