Conciseness

Nearly all of the body is project-specific knowledge Claude cannot already know (pilot-run artifacts, the runtime's 20-dispatch concurrency cap, the ~5-20x whole-dispatch overhead measurement artifact, scope cost math), and the two heredoc scripts are lean. It is not a 5 because several passages carry justificatory padding that could be trimmed — "This is load-bearing, not a style preference", "visible, not assumed away", and the multi-sentence rationales after "Never fabricate a token count" and "a made-up accuracy figure is worse than a visibly incomplete one, because it's indistinguishable from a real one in the report". It stays above anchor 3 because none of it explains concepts Claude already knows — every long passage encodes a pilot-learned failure mode.

Actionability

Two copy-paste-ready executable blocks (the variant-generation python heredoc and the metrics-computation python heredoc), an exact validation command (`uv run --no-project --with jsonschema scripts/validate_blueprint.py -`), exact output paths (`$SCRATCH/results/composition/<variant-name>/<task-id>.json`), and a per-row recipe for the accuracy sweep. It misses anchor 5 only because the sub-agent dispatch instructions in Step 2 are described precisely but never given as a verbatim prompt template, and `SCRATCH=<a scratch directory — never a path inside the repo checkout>` is a placeholder rather than a concrete default — minor gaps, not the missing-steps pattern of anchor 3.

Workflow Clarity

Five clearly sequenced steps with explicit validation and feedback loops throughout: a pre-flight scope confirmation with computed cost counts before the expensive phase; schema validation of every blueprint with a re-emit-once retry loop ("if it doesn't validate, ask that one sub-agent to re-emit once, quoting the validator's error"); a byte-identical check of the all-four variant delegated to `tests/model_right_sizer/test_ablation_layers.py`; a stall-diagnosis and recompute-from-filesystem recovery procedure for batch dispatches; and an explicit "don't infer 'still running' purely from the absence of a failure" checkpoint. This matches the anchor-5 pattern of validate → fix → retry with recovery loops for a batch operation.

Progressive Disclosure

No bundle files exist under references/, scripts/, or assets/ — the body is the whole skill, with all methodology pushed one level out to clearly signaled repo files (`../../eval/ablation/DESIGN.md` — "read that file first; this SKILL.md is the runbook, not the methodology essay" — plus layers.py, benchmark_tasks.json, metrics.py, and a Related section). The body self-describes as a runbook and appropriately defers detail rather than inlining it. It misses anchor 5 because some long inline digressions (the ~40-line stall-recovery procedure and the dispatch-overhead normalization discussion) read as reference material that could be split into their own linked file, and none of the referenced external files are part of this bundle, so a reader outside the parent repo cannot follow them.

Specificity

The description enumerates the full concrete pipeline: "Renders layer-ablated variants (any of the 16 layer subsets), runs a fixed six-task benchmark through each variant's Pass A blueprint", "actually executes the recommended build and scores whether real effort stayed within the predicted budget (wrapping `classify_budget_adherence`)", and "Reports each layer's effect in ISOLATION vs. a zero-layer baseline, and every COMBINATION across the full 16-subset grid". Coverage spans render, benchmark, execute, score, and report with no gaps — a clear anchor-5 match, not anchor 4's 'minor gaps in coverage'.

Completeness

Both halves are explicit: the 'what' is the measured ablation study described in operational detail, and the 'when' is a literal "Use when someone says ..." clause with four concrete trigger phrases. This matches the anchor-5 example structure exactly (what + 'Use when ...' with concrete triggers), and is clearly above anchor 4's 'when could be more explicit'.

Trigger Term Quality

Four natural quoted trigger phrasings are given ("does the Token Economics layer actually change anything", "ablate the research layers", "run the layer-ablation study", "audit model-right-sizer's citations empirically") — genuinely what a user would say. Not anchor 5 because coverage is strong but not exhaustive: a few plausible natural variants (e.g. "which citation layers matter", "is the BudgetThinker layer real") are absent; not anchor 3 because the quoted phrases already capture the common phrasings users would use for this niche.

Distinctiveness Conflict Risk

It occupies a clear niche — an empirical ablation study of one named agent's (model-right-sizer's) four named citation layers — with distinct triggers unlikely to fire for any other skill. The named layers (Token Economics, IBPO, BudgetThinker, Speculative Decoding) and the wrapping of `classify_budget_adherence` make confusion with a generic benchmarking or sizing skill essentially impossible.