Conciseness

The bulk is dense, harness-specific instruction Claude cannot infer (evaluator discovery rules, annotation fan-out protocol, triage categories, YAML gotchas), but there are several passages of rationale that could be trimmed: the context-rot explanation in step 6, 'Why this matters:' after the agreement rule, and meta-commentary like 'This frames the task as a goal with constraints, not a procedural recipe. Cross-model fan-out is only useful if...'. Not 2 because almost none of it re-teaches concepts Claude already knows (no 'what is YAML' padding); not 4 because the repeated just-it-yourself-rationale blocks appear in at least three places and read as over-explanation of the author's design choices.

Actionability

Fully executable throughout: exact commands ('uv run python -c "from evals.pxi.harness.datasets import load_dataset; load_dataset(\'<name>\')"', 'uv run python -m evals.pxi.harness.run_experiment --dataset <name>'), a per-model invocation table, a working bash example for the Codex wrapper, concrete YAML blocks for variant lists and metadata, and specific file paths for every artifact. Not 4 because the concrete examples cover the common cases end-to-end with nothing left pseudocode.

Workflow Clarity

A numbered 7-step workflow with explicit validation checkpoints (step 7: 'Parse + validate schema', 'Run the experiment end-to-end and triage every failure before calling the dataset done'), a three-bucket failure triage table, a fix-order rule ('harness/evaluator issues first, dataset issues second'), and a re-run loop ('Re-run the full experiment after each batch of fixes. Stop when...'). This is a batch operation and validation is present, so the cap does not apply. Not 4 because error-recovery feedback loops and checklists (coverage dimensions in step 4) are explicit, not implied.

Progressive Disclosure

Good structure: clearly headed sections, a verified one-level-deep script reference ('scripts/annotate_via_codex.sh' exists and its flags in the body match the file), and pointers to live repo files instead of duplicated schemas ('Do not copy a schema from this skill into prompts'). Not 5 because some content that would sit better in a reference file is inlined in SKILL.md — notably the verbatim ~40-line annotation task-instruction prompt template and the evaluator matching-semantics reference — making the body a ~575-line near-monolith; not 3 because what is inlined is well-signaled with headers and the only bundle reference is real and clearly linked.

Specificity

Multiple concrete actions are named: 'Generate synthetic evaluation datasets for the PXI eval harness (evals/pxi/)', 'Inspects whichever evaluators currently live under evals/pxi/evaluators/', and 'pauses to recommend a new evaluator'. Coverage is comprehensive for the skill's niche. Not 4 because no meaningful action within scope is left out — even the boundary behavior (recommending a new evaluator on a gap) is explicit.

Completeness

Explicitly answers both questions: the 'what' is 'Generate synthetic evaluation datasets for the PXI eval harness (evals/pxi/)' and the 'when' is 'Use whenever the user asks to create, author, draft, expand, or audit an eval dataset... including phrases like...'. Both are concrete and paired in one tight description. Not 4 because the 'when' clause is fully explicit with trigger phrases, not merely present.

Trigger Term Quality

Quotes natural user phrasings users would actually say: 'write evals for <tool>', 'test PXI behavior', 'synthetic dataset for PXI', 'cover this tool with eval examples', 'find gaps in our PXI eval coverage', plus verb synonyms (create, author, draft, expand, audit) and the concrete path trigger evals/pxi/. Not 4 because both synonyms and concrete phrase-level triggers are present, matching the anchor-5 pattern of PDF files/PDFs/forms/.pdf.

Distinctiveness Conflict Risk

The niche is sharply delimited — PXI eval datasets under evals/pxi/ with named evaluator directories — so it is unlikely to fire for adjacent skills (general eval writing, dataset tooling outside PXI). Not 4 because every trigger phrase is PXI-qualified, leaving minimal overlap risk with closely related skills.