Content
92%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.
A strong, expert-level skill body: an explicitly sequenced eight-step workflow with validation and feedback loops at every risky stage, copy-paste-ready commands and flags, and clean one-level-deep offloading to nine real reference files. Its only weakness is length — the cross-reference list and five worked examples could be tightened without losing substance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence — no primers on what fuzzing or sanitizers are, and nearly every line is non-obvious domain judgment ("Dedupe by crash type plus the top three frames in project code", "Forty crash files are often two bugs", the Apple-clang/nightly/Windows toolchain gotchas). A few sections could be trimmed: the 10-skill cross-reference paragraph, the five worked examples (~40 lines where two would carry the patterns), and parts of "Not for" that repeat the Boundaries section; this places it at efficient-with-minor-trimmings rather than the every-token-earns-its-place top anchor. | 4 / 5 |
Actionability | Guidance is copy-paste ready throughout: executable rg/grep inventory commands with a portable fallback, exact sanitizer and limit flags ("-fsanitize=fuzzer,address,undefined", "-rss_limit_mb=2048", "-max_total_time=600"), a per-language engine table, a fully specified ext-keys schema, and concrete validation commands ("node references/validate-findings.mjs findings.json"). The common cases are covered with specifics down to file paths and output directories; no pseudocode stands in for executable content. | 5 / 5 |
Workflow Clarity | The eight-step workflow (Inventory → rank → harness → seeds → run → triage → verify → regression/CI) is explicitly sequenced with validation checkpoints and feedback loops throughout: reproduce each crash before minimizing, "Flaky: check for leaked state or threads before going on", harness-bug detection ("A crash caused by the harness breaking the contract is a harness bug: fix it and rerun"), the refutation pass before any verdict, "Write findings.json first and validate it", and "Verify the fix" in the report template. Because this batch/run-heavy workflow has explicit validation at every stage, the destructive/batch cap at 3 does not apply. | 5 / 5 |
Progressive Disclosure | SKILL.md is an overview that pushes per-language templates, build commands, the triage class table, severity calibration, verification procedure, findings format, and CI setup into nine reference files, each present in references/ and each introduced with a clear signal of what it contains (e.g. "Per-language templates, structured-input helpers and build commands are in [references/harness-templates.md]"). All referenced paths resolve to real bundle files exactly one level deep with no nested reference chains, matching the clear-overview anchor. | 5 / 5 |
Total | 19 / 20 Passed |