Content
80%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is a strong, actionable red-team reference: dense, copy-paste-ready attack strings and a runnable probe config with no padding. Its main gaps are the absence of an explicit run-detect-validate workflow and any progressive disclosure into separate reference files.
Suggestions
Add a short numbered workflow with validation checkpoints, e.g. 1) run the probe config, 2) scan outputs against the Detection signals, 3) rate via the Severity table, 4) re-run with the next strategy only if prior signals are negative.
Consider splitting the seven technique details into a references/ file (e.g. TECHNIQUES.md) and keeping SKILL.md as an overview pointing to it, to improve progressive disclosure for this longer skill.
Add an explicit feedback loop for the batch probe step (e.g., 'if a detection signal fires, confirm with a second probe strategy before classifying severity').
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence — it never explains basics like what a system prompt is — and each example earns its place (e.g., the 10 extraction prompts 'in order of subtlety'), matching the lean-and-efficient anchor. | 3 / 3 |
Actionability | It supplies literal copy-paste attack strings and a ready-to-run YAML probe config with concrete plugin ids and strategy names, meeting the 'fully executable, copy-paste ready' anchor. | 3 / 3 |
Workflow Clarity | A loose ordering exists ('try in order of subtlety'; Techniques → Probe → Detection → Severity → Defender) but there are no explicit validation checkpoints or feedback loops for the batch probe operation, which the scoring notes cap at 2. | 2 / 3 |
Progressive Disclosure | The skill is a single self-contained file with well-organized sections, but at ~115 lines with seven full techniques inline and no external references, the under-50-line 'well-organized sections → 3' carve-out does not apply; content that could be split stays inline. | 2 / 3 |
Total | 10 / 12 Passed |