Content
42%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body contains an unusually rigorous, internally consistent behavioral contract with a real, well-signaled reference file, but it buries its actionable core in a ~500-line Lean formalization that restates the same protocol in four registers. A model reading it gets precise map and gate semantics yet no worked example, and the token cost is enormous relative to the behavioral payload.
Suggestions
Move the Lean formal block to a reference file (e.g. references/formal-contract.md) and keep SKILL.md to a concise prose protocol; the same rules currently appear in the Definition prose, FLOW, MORPHISM, and TOOL GROUNDING comments, and one passage repeats verbatim.
Add one worked example — a short sample map with its gate, a person's correction, and the closing relay/handoff turn — so the contract's execution is concrete rather than inferred from axioms.
Restate the operational sequence as a short numbered checklist with explicit checkpoints (observe pointer and inventory → draft whole-method map → present → gate on correction → relay and hand off in the closing turn → stop), keeping the formal detail behind the reference.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is ~540 lines dominated by a Lean 4 formal block in which the same content is restated four times over — prose Definition, the FLOW comment, the MORPHISM comment, the TOOL GROUNDING strings, plus a verbatim repetition of "Once it has run, the person can say: this work proceeds this way..." in both the Definition and the ConductedMethod doc comment. The actionable core (draft a whole-method map, show it, gate on the person's correction, relay and hand off) could be carried in a fraction of the tokens. This fits anchor 2 (noticeably verbose, several padded sections); it avoids anchor 1 only because the material is an original formalism rather than explanations of concepts Claude already knows. | 2 / 5 |
Actionability | There is genuinely specific guidance — the `.map` grounding string enumerates exactly what the map must show, `ConductedMethod` lists the handoff fields, and the gate semantics are pinned ("what the map got wrong... or take the method; silence holds and takes nothing") — but there are no worked examples (no sample map, no example taking or handoff turn), and the semantics of every key predicate are deliberately delegated to unstated "your judgment" axioms. This matches anchor 3 (some concrete guidance but incomplete, missing key details); not 4 because a model could not execute this without substantial inference the skill never grounds, and not 2 because the map contents and gating rules are far more concrete than high-level hints. | 3 / 5 |
Workflow Clarity | The sequence is present and coherent — start, observe (pointer + inventory), first map, per-utterance gating, relay/handoff on a taking, then stop — laid out in both the FLOW comment and the `conduct`/`start` definitions, with failure handling specified ("a pointer that does not resolve... is a finding the map shows, with what was tried; it closes nothing"). However, the checkpoints are formal and implicit rather than operational: no explicit verify-then-proceed step, and the flow must be reconstructed from dense Lean recursion and cross-referenced comments. Fits anchor 3 (sequence present but checkpoints missing or implicit); not 4 because the density leaves practical checkpoints and error-recovery loops implicit. This is a dialogue protocol, not a destructive or batch operation, so the workflow-clarity cap does not apply. | 3 / 5 |
Progressive Disclosure | The bundle structure is real and well-used: `references/round-composition.md` exists, is one level deep, and is clearly signaled with occasion-bound load conditions ("Read references/round-composition.md when terminology or wording must persist, content belongs to another turn, or placement relative to the gate is in question"). But the SKILL.md body itself inlines a ~500-line formal block that is precisely the material progressive disclosure would move to a reference, leaving the main file a monolith. Fits anchor 3 (some structure, references present and signaled, but content that should be separate is inline); not 2 because section headers and a properly signaled reference do exist, and not 4 because the dominant content of the file is the inlined block. | 3 / 5 |
Total | 11 / 20 Passed |