Conciseness

The body is lean with no padding and no explanation of concepts Claude already knows — every line is non-obvious judgment guidance ('A client passing a case does not prove the harness caused the behavior', 'Unavailable providers are not passes'). It assumes Claude's competence and each directive earns its tokens, matching the top anchor rather than the 4 anchor's 'minor instances of over-explanation'.

Actionability

There is some concrete guidance — 'inspect evals/promptfoo/promptfooconfig.yaml', 'make install-promptfoo', 'make eval-promptfoo', 'pin the evaluator and commit the dependency lockfile' — but it is incomplete for the core task: no example promptfoo case, assertion syntax, or config snippet showing how to actually author an evaluation. It is not a 4 because the central 'build a case and grader' section stays at the level of principles ('Prefer deterministic assertions for structure, exact facts, and prohibited actions') without a concrete worked example, and not a 2 because commands and file paths do give real, executable hooks.

Workflow Clarity

A clear define → build → run-and-interpret sequence with most checkpoints present: 'Exercise the grader with a known correct response and plausible incorrect responses before trusting it', 'check its judgments against examples reviewed by a person', and 'Report behavioral failures separately from provider errors, timeouts, missing credentials'. It falls short of the 5 anchor because error-recovery feedback loops (what to do when a run fails or flakes) are implied rather than explicit, and the steps are prose sections rather than an explicit ordered sequence.

Progressive Disclosure

No bundle files exist (references/, scripts/, assets/ are absent), and the skill needs none: the ~68-line body is self-contained, cleanly sectioned, and everything it contains is core guidance that belongs in SKILL.md. The one file pointer it gives (the Arnesto config path and make targets) is clearly signaled and one level deep, matching the well-organized-sections standard for simple self-contained skills.

Specificity

Names concrete actions — 'Design and run promptfoo evaluations', 'comparisons across prompts, agent rules, and models', 'explicit assertions' — with only minor gaps in coverage (e.g., no mention of report generation or CI integration). It is not a 5 because it stops short of the comprehensive multi-action coverage of the top anchor, and not a 3 because it lists several specific actions rather than just one or two.

Completeness

Explicitly answers both questions: what ('Design and run promptfoo evaluations for LLM behavior regressions or comparisons across prompts, agent rules, and models') and when ('Use for repeatable behavioral cases with explicit assertions, not ordinary unit tests or static configuration checks'). The 'Use for' clause includes concrete trigger phrasing plus negative boundaries, matching the top anchor; a 4 would require the 'when' to be less explicit, which it is not.

Trigger Term Quality

Good natural keyword coverage: 'promptfoo evaluations', 'LLM behavior regressions', 'prompts', 'agent rules', 'models', 'assertions', 'unit tests'. A few natural phrasings users would say are missing, such as 'evals', 'evaluate prompts', or 'A/B test prompts', keeping it below the 5 anchor's synonym-and-variation coverage.

Distinctiveness Conflict Risk

'promptfoo evaluations' carves out a clear niche with distinct triggers, and the explicit exclusion ('not ordinary unit tests or static configuration checks') prevents conflicts with general testing and linting skills. Minimal overlap risk with closely related skills like generic test-writing.