Content
63%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, dense, highly executable debugging guide whose real value is the non-obvious material: Hermes-specific attach flows, sourcemap/child-process/port pitfalls, and the verification checklist. Its weaknesses are structural: reference-grade content (REPL table, CDP driver, profiler recipes) is inlined in a ~300-line single file with no external references, and the REPL table plus one-shot recipes restate knowledge Claude already has.
Suggestions
Move the REPL command table, the full CDP driver script, and the heap/CPU-profile recipes into references/ files (e.g. references/repl-commands.md, references/cdp-driver.js) and keep SKILL.md as a lean overview with one-level-deep pointers.
Delete or compress the One-Shot Recipes section — it restates flows already covered by Quick Reference and Attaching sections — and trim the REPL table to only non-obvious commands (sb, repl, exec, watchers).
Fix the non-executable snippet detail: writeFileSync('~/...') does not expand '~' in Node — use os.homedir() or a literal path so the profile/snapshot recipes are copy-paste ready.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly lean commands and code with high-value non-obvious material (Hermes-specific attach flows, pitfalls, sourcemap warnings), but it includes unnecessary known-material padding: the ~25-line `debug>` REPL command table restates standard Node-documented commands Claude already knows, and the One-Shot Recipes section largely restates earlier flows. This fits 'mostly efficient but includes some unnecessary explanation or could be tightened' — not anchor 4, since the redundant material is more than minor. | 3 / 5 |
Actionability | Nearly everything is concrete and executable: copy-paste bash (SIGUSR1 attach, pgrep/curl/jq WS-URL lookup, vitest single-file debug) and a complete CDP driver script covering the common scenarios. It falls short of fully copy-paste-ready anchor 5 on small gaps: `require('fs').writeFileSync('~/.hermes/...')` will not expand the tilde in Node fs, and `<uuid>`/`<hermes-agent-repo>` placeholders need manual substitution — 'concrete code or commands with minor gaps', i.e. anchor 4. | 4 / 5 |
Workflow Clarity | Multi-step processes are clearly sequenced (numbered launch → SIGUSR1 → WS-URL → attach; build → inspect-brk → sb/cont → repl), and the Verification Checklist plus pitfall-driven recovery ('if the first breakpoint doesn't hit, you likely missed --inspect-brk') provide real validation and feedback loops. Not anchor 5 because validation sits in an end-of-file checklist rather than being integrated into each workflow, and the vitest/heap-snapshot workflows carry no verification steps; no destructive/batch cap applies since operations are read-only debugging. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, assets/); the entire skill is a single ~300-line file. Internal section headers are good, but reference-grade material that clearly belongs in separate files — the REPL command table, the full CDP driver script, and the heap/CPU-profile recipes — is inlined in SKILL.md. This matches 'some structure but... content that should be separate is inline' (anchor 3); it is not anchor 2 since the sections are well-organized and navigable, and the under-50-line exception does not apply. | 3 / 5 |
Total | 14 / 20 Passed |