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 well-engineered operational protocol: phased workflow, executable file-type handlers, explicit safety gating, and honest failure handling. The main costs are repetition of the safety gate and test-facing boilerplate sections, plus a monolithic single-file layout whose long templates belong in reference files.
Suggestions
State the scan_paths safety gate once (its dedicated section) and reference it from Phase 1 and Anti-Patterns instead of restating it verbatim.
Move the manifest template and brain-page format into references/ files (e.g. references/manifest-template.md) and link to them, keeping SKILL.md as an overview.
Delete or merge the 'Contract' and 'Output Format' conformance-test stubs — content written for a test harness, not for the executing agent, spends tokens without instructing.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient and assumes Claude's competence (no explanations of what mbox or PST formats are), but it is padded in places: the safety gate is restated four times (dedicated section, Phase 1 step 1, 'When it fails', 'Anti-Patterns'), and the 'Contract'/'Output Format' sections openly exist 'for the conformance test' rather than to instruct the agent — the 'could be tightened' anchor, not the 'minor instances' of 4. | 3 / 5 |
Actionability | Concrete, executable guidance dominates: runnable Python for .mbox and .docx, a .pst magic-byte validation check, bash commands, an exact gbrain.yml config, and full page/manifest templates. It stays at 4 rather than 5 because the advertised 'dropbox' and 'backblaze' source types have no corresponding code or commands in the file-type handlers. | 4 / 5 |
Workflow Clarity | The three phases give a clear numbered sequence with real checkpoints for a batch operation: the scan_paths gate runs first, 'Never re-show — check the manifest before presenting anything', PST validation before extraction, and error-recovery loops (poll 'gbrain write-request', merge on 'revision_conflict'). It falls short of 5 because recovery guidance lives in a separate 'When it fails' section rather than inline as validate→fix→retry steps within the crawl workflow. | 4 / 5 |
Progressive Disclosure | Headers organize the 310-line body well and the external convention links (../conventions/quality.md, ../_brain-filing-rules.md) are one level deep, but there is no references/ bundle at all: the ~45-line manifest template and the brain-page format are inlined monolithically where they clearly belong in separate files — the 'content that should be separate is inline' anchor 3 rather than the 'minor organization gaps' of 4. | 3 / 5 |
Total | 14 / 20 Passed |