Content
85%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 is an exceptionally well-engineered behavioral spec: concrete file contracts, exact commands, deterministic branching for every failure mode, and a cut-off recovery protocol. Its one weakness is redundancy — the write-to-notes-as-you-go imperative and its rationale are repeated several times across the task section and could be consolidated to roughly two statements without losing any behavioral force.
Suggestions
Consolidate the repeated write-as-you-go injunction into one normative rule plus one sentence of rationale — "write every settled step to DIR/notes.md before reasoning further; unwritten reasoning is lost on cut-off; keep going after each write" — instead of restating it four times across The task section.
State the output-allowance/cut-off rationale once rather than twice ("a message that is cut off is normally followed by..." and "each message's output allowance also covers your private reasoning" overlap).
The closing reply contract (where proof.md is, and whether it resolves the problem) could move into the Setup section's file-contract list, letting The task section end at the proof.md write.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The Arguments and Setup sections are dense but every clause handles a distinct edge case (DIR parsing, bad setting tokens, file-path detection, resume states) — not padding. However, the write-as-you-go rule is restated roughly four times ("write your work product out as you go", "whenever you settle something... write it down before reasoning further", "Writing to the notes is not finishing — keep going after each write", "Many short written steps beat one long unwritten one") and the cut-off rationale is explained twice. Mostly efficient but noticeably tighter possible — anchor 3, not 4 because the repetition goes beyond minor over-explanation. | 3 / 5 |
Actionability | Fully actionable without code: exact paths (DIR/problem.md, DIR/notes.md, DIR/proof.md), exact commands (`mkdir -p`, `cp`, `cmp`), the default DIR ("./math-proof-solo"), verbatim one-sentence error replies for every failure branch, and a concrete resume protocol. An instruction-only skill whose guidance is copy-paste executable — anchor 5. | 5 / 5 |
Workflow Clarity | The sequence is unambiguous (parse arguments → setup/resume check → staged reasoning with mandatory notes writes → write self-contained proof.md → brief reply) and every branch has an explicit checkpoint and recovery path: `cmp` to detect a different problem, detecting an already-completed run, "If a message of yours is cut off, re-read DIR/notes.md and continue from the last thing written there; never start over", and the missing-input usage reply. Explicit validation and error-recovery feedback loops — anchor 5. | 5 / 5 |
Progressive Disclosure | There are no bundle files (no references/, scripts/, or assets/ exist) and nothing in the ~60-line body warrants a separate file for this single-task skill. Content is organized into clearly labeled sections (Arguments, Setup, The task) with the deliverable's contract stated up front. Under the simple-skill exception, well-organized sections with no nested references score 5. | 5 / 5 |
Total | 18 / 20 Passed |