Conciseness

Quotes: the contamination incident is narrated at length ('A first attempt... generated all three "blind" draws inline... it's transcribing the answer key and then reporting that it correlates with the answer') and then restated in step 2 ('Do not generate the draws yourself inline, even carefully, even if you believe you can reason about it "as if blind"') and again in the NOT-do section ('It does not accept an inline, self-authored "blind" draw under any framing — see the incident above'). The dilution lesson likewise appears in the intro, step 4, and step 5. The content is project-specific, not concepts Claude already knows, so it is not a 2, but the same critical rules are stated three times where once plus a pointer to the linked incident write-up would carry the same force.

Actionability

Quotes: 'Dispatch 3 or more separate Task/Agent calls, each given only: the full definition + 0.0/1.0 anchors for every signal being rated', 'average each signal's value across the draws. Compute... the coefficient of variation (stdev / mean across the draws)', 'the candidate signal alone; the existing signal sum... plus the candidate at equal weight', and 'Add regression tests mirroring tests/model_right_sizer/test_token_ceiling_formula.py's existing test_*_dilutes_the_existing_signal_correlation pattern'. This is concrete, executable guidance with exact file paths, thresholds, and naming patterns — as an instruction-only skill, absence of code is not penalized per the scoring notes. It is not a 5 because there is no copy-paste-ready dispatch prompt template (the single most fragile step) and no command showing how the correlation is actually computed (script call or function name).

Workflow Clarity

Quotes: an 8-step numbered sequence from 'Pick the candidate signal and a held-out task' through 'A weight change is a proposed diff for a human to review', with explicit validation checkpoints throughout: CV as 'the noise-magnitude diagnostic, not a pass/fail gate', the combined-sum correlation bar ('Judge a candidate on the combined-sum delta, not the standalone figure'), the replication gate ('Only propose... once the SAME candidate clears the correlation bar on a second, different held-out task'), and error-recovery paths ('if no fresh task exists yet, say so explicitly and treat the result as weaker evidence'; 'If the runtime has no sub-agent dispatch available at all, stop and say so'). This matches the anchor-5 shape: clear sequence, explicit validation, feedback loops for failure modes.

Progressive Disclosure

The body is organized into clear sections (incident, 'What to do' steps, 'What this does NOT do', 'Related') and the Related section lists well-signaled one-level-deep references with markdown links, e.g. '[`../../eval/token_ceiling_formula.py`]... — the module under test', '[`../../eval/tuning/overfitting_guard.py`]... — HOLDOUT_TASKS'. No nested-reference chains exist, and no bundle files are provided to score against. It is not a 5 because the ~30-line incident narrative largely duplicates the linked results write-up ('2026-08-22-second-signal-experiment-genuinely-blind.md') and could live in a reference file with a two-line summary in the body.

Specificity

Quotes: 'Test whether a candidate real-work signal... deserves a nonzero default weight', 'Dispatches 3+ independent draws per held-out task, computes per-signal CV and Pearson correlation (alone, and added to the signal sum...)', 'requiring replication on a SECOND held-out task before proposing a nonzero weight'. These are multiple concrete, comprehensive actions covering the full workflow (test, dispatch, compute CV/correlation alone and combined, replicate, propose weight). It is not a 4 because there are no coverage gaps — every step of the method is named concretely; the explanatory asides ('which would turn blind rating into transcribing the answer key') are constraints, not vagueness.

Completeness

Quotes: 'Test whether a candidate real-work signal... deserves a nonzero default weight, via the blind multi-draw rating + correlation methodology' (clear what) and 'Use when someone says "test this new signal"...' (explicit when with concrete trigger phrases). Both what and when are answered explicitly and concretely, matching the anchor-5 example pattern; a 4 would require the 'when' to be merely present but less specific, which is not the case here.

Trigger Term Quality

Quotes: 'Use when someone says "test this new signal", "does [signal] deserve a nonzero weight", "re-run the signal validation experiment", or "validate real-work signals against real data"'. Four natural user phrasings is good keyword coverage for this niche domain. It is not a 5 because common variations like 'should this signal keep a zero weight', 'check the correlation', or references to the sibling tuning skills are absent, leaving a few natural trigger phrasings uncovered.

Distinctiveness Conflict Risk

Quotes: 'Test whether a candidate real-work signal in `eval/token_ceiling_formula.py`... deserves a nonzero default weight' and trigger phrases like 'validate real-work signals against real data'. This is a clear niche (signal weight validation for one specific formula module) with triggers that would not fire for the sibling skills (holdout tuning, prompt tuning) or any generic evaluation request. Minimal conflict risk.