Content
71%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-structured routing skill: concrete probes, exact sub-skill paths, and an explicit pivot/pacing discipline give it strong actionability and workflow clarity, and the overview-to-sub-skill split is textbook progressive disclosure. Its main weakness is token weight — many routing rows duplicate sub-skill-level technique detail that the router does not need to make a load decision, and a few conventions (PIVOT.md, SUMMARY.md) are referenced without definition.
Suggestions
Trim the 'Covers' column of the routing table to 2-3 word technique labels; move payload-class detail (e.g. blind-sqli's probe-loop specifics, smuggling's CL.TE/TE.CL variants) into the sub-skills, keeping only the 'When to Load' signal in this file.
Define or point to the `exploit/PIVOT.md` and recon `SUMMARY.md` conventions (one line each) so the pivot rule is self-contained rather than assuming external context.
Extend the Quick Detection probes to cover the higher-frequency routed classes with no probe (e.g. IDOR/object-ID check, JWT bearer-token check) so 'Unknown → run probes' is more likely to resolve the route.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The routing table, decision flow, and pacing rules are mostly high-signal, but many table rows inline sub-skill detail that belongs in the sub-skills themselves (e.g. blind-sqli's "token-fingerprinting probe loops, arithmetic-multiplication boolean evaluation, hex-encoded literals, exponential-probe binary search"; similarly dense smuggling and crypto rows). Mostly efficient but could be tightened to a leaner when-to-load signal per row. | 3 / 5 |
Actionability | The Quick Detection section gives eight copy-paste-ready curl probes and every table row carries an exact `load_skill("...")` path, which is concrete and executable. Minor gaps remain: probes use a `<TARGET>` placeholder and the pacing rule references `exploit/PIVOT.md` and "recon SUMMARY.md" conventions that are never defined or linked here. | 4 / 5 |
Workflow Clarity | A clear sequence exists (detect via probes → decision flow → load sub-skill), and the Payload Pacing rule supplies explicit feedback thresholds (3 same-class negatives → next bypass class; 3 exhausted classes → pivot) that function as validation checkpoints. Verification of exploitation success is delegated to sub-skills and the referenced system-prompt agent loop rather than restated, leaving a minor checkpoint gap at the router level. | 4 / 5 |
Progressive Disclosure | This is a model router skill: the body is a pure overview whose only depth is one-level-deep, clearly signaled `load_skill` pointers into specialized sub-skills, organized in a scannable table plus a decision tree. No bundle files (references/, scripts/, assets/) exist and none are implied, and the split between overview and detail is exactly where it should be. | 5 / 5 |
Total | 16 / 20 Passed |