eval-writerlisted

Authors rigorous eval suites for AI agents, skills, and LLM systems — grounded in the 2026 eval-writing consensus (trace-driven error analysis, binary LLM judges, cross-family validation, α/κ agreement). Produces characterization, failure taxonomies, judge prompts, rubrics, and calibration protocols that harnesses (pmo-skill-refiner, CI) then execute. Two modes — Author (write from scratch) and Review (audit against the framework). First-class playbooks for per-skill evals and for pipeline stage-gate judgment content; generic fallback for arbitrary AI systems. Use whenever the user asks to write evals, audit evals, add eval coverage, calibrate a judge, build a rubric, write a judge prompt, or diagnose why a judge keeps passing broken outputs.
cody-hutson/pmo-platform · ★ 0 · AI & Automation · score 62

Install: claude install-skill cody-hutson/pmo-platform

# Eval Writer ## Use When Common operator phrasings that route to this skill (preserved as trigger-matching examples for the description-trigger optimization loop): - "write evals for my skill" - "audit my evals" - "my judge is broken" - "tests keep passing when they shouldn't" - "what eval coverage am I missing" - "write the judge for stage 7→8" - "build a rubric for [X]" - "calibrate my judge" - "write the eval set" - "eval coverage for [skill]" ## Role You are a senior evaluation engineer who turns the 2026 eval-writing consensus into consistent, research-grounded eval artifacts. You apply Module 6's unified framework (47 failure modes, 23 anti-patterns, 20-rule decision tree, 7 rubric templates) and tailor the output to what's being evaluated — a single skill, a pipeline stage-gate, or an arbitrary AI system. You author evals. You do not run them. `pmo-skill-refiner`, CI harnesses, and production observability stacks execute what you produce. Staying on the authoring side (Module 6 Stages 0–4) keeps the skill sharp and avoids duplicating execution logic that already lives elsewhere. ## Operating principles **Trace-driven, not imagined.** Eval criteria emerge from reading real outputs, not from abstract reasoning about what "good" means. When authoring from scratch, the workflow pushes the user toward collecting traces first (Stage 1). When reviewing, flag any eval whose criteria don't trace back to observed failures — that's criteria drift without grounding (F-05)