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.
An exceptionally disciplined operational skill: concrete commands, consent-gated destructive edits with validation and revert paths at every gate, and near-zero token waste. The main weakness is monolithic structure — everything lives inline in one headerless file instead of splitting conditionally-needed sink-class procedures into referenced files.
Suggestions
Add markdown section headers (e.g., ## Sinks, ## Fix procedures by class, ## Reporting savings, ## Binding rules) so the file can be navigated without reading it end to end.
Move the per-sink-class procedures (RECURRING_CONTEXT, SKILL_DISTILLATION, MEMORY_HEALTH) into one-level-deep reference files (e.g., references/sink-classes.md), keeping SKILL.md as the overview plus the REDUCIBLE flow and binding rules; they are only needed once the user selects a sink of that class.
Consolidate the repeated guard phrases ('never make the agent dumber', 'never present as verified') into the single 'Binding rules' section and reference it from each procedure to trim the duplicated caveats.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense, imperative rules with no padding and no explanation of concepts Claude already knows. A few binding guard phrases repeat ('never make the agent dumber' appears in both the preamble and the RECURRING_CONTEXT gate; the 'never call it verified' rule appears twice with different wording), which could be consolidated into the single binding-rules section. | 4 / 5 |
Actionability | Fully executable, copy-paste-ready commands with concrete flags throughout — 'caveman learn report --json', 'caveman learn apply <sink_id> --dry-run', 'caveman mem remember --stdin < <temp file>', 'caveman learn experiment start <label> --sink <sink_id> --fix-kind skill_distillation' — plus an exact re-verification procedure (re-segment the turn, compare sha256) for the locator check. | 5 / 5 |
Workflow Clarity | Each sink class has a clearly sequenced procedure with explicit validation gates and error-recovery loops: dry-run before apply, the net-token-negative gate ('if after is not below before, revert and report'), sha256 mismatch abort, the recall-path check that forces a revert if recall returns nothing, and a closing binding-rules checklist. This satisfies the destructive/batch validation requirement fully. | 5 / 5 |
Progressive Disclosure | The skill is a single ~170-line file with no bundle files and no markdown section headers — only ALL-CAPS prose labels. The per-sink-class procedures (RECURRING_CONTEXT, SKILL_DISTILLATION, MEMORY_HEALTH) are needed only when the user picks that sink, so they are prime candidates for one-level-deep reference files, yet all are inlined. Structure exists (class labels, binding-rules section) but the file is over the 50-line threshold where sections alone would earn a 5. | 3 / 5 |
Total | 17 / 20 Passed |