← ClaudeAtlas

rubric-panellisted

Score one or more code artifacts against an anchored rubric using a persona-diverse judge panel; produces dimension scores plus findings classified as defect, noise, spec-hole, or undeclared-commitment. Use when asked to "score this implementation", "judge these two solutions", "rate this module against a rubric", or compare competing implementations of the same spec, and when invoked by layer-cycle. For choosing between technologies, frameworks, or vendors, use tech-eval-research instead.
hsb3/dotfiles-agents · ★ 0 · AI & Automation · score 70
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