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.
The body is a highly actionable, well-structured review playbook with concrete vulnerable/safe pattern pairs and a clear confidence-gated workflow. Its two weaknesses are redundancy (the attacker-vs-server-controlled guidance is repeated three times) and a progressive-disclosure defect: the language and infrastructure guides it tells the agent to load are missing from the bundle entirely.
Suggestions
Ship the referenced languages/*.md and infrastructure/*.md files, or remove the "Load Language Guide" and "Load Infrastructure Guide" steps and fold any essential guidance into the existing references.
Consolidate the three repetitions of the attacker-controlled vs. server-controlled material (Server-Controlled Values table, Verify Exploitability table, and the overlapping Quick Patterns examples) into a single section.
Add an explicit loop-back rule when research downgrades confidence (e.g. "if research reveals server control or upstream validation, move the finding to Needs Verification or drop it").
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly dense, high-signal tables, but there is noticeable duplication: the attacker-controlled vs. server-controlled distinction appears three times (the "Server-Controlled Values" table, the "Verify Exploitability" table, and the SSRF/path-traversal examples in "Quick Patterns"), and "Research Before Flagging" repeats the same bullets already stated in "Scope: Research vs. Reporting". This matches anchor 3 (mostly efficient but could be tightened) rather than anchor 4, where only minor trimming would be needed. | 3 / 5 |
Actionability | Guidance is fully concrete and executable for a review skill: exact vulnerable and safe code patterns (e.g. "requests.get(request.GET['url'])" vs "requests.get(settings.API_URL)"), specific framework-mitigation rules ("{{ var|safe }}", "dangerouslySetInnerHTML", ".raw()"), a severity table, and a copy-paste-ready output template. It matches anchor 5 — specific examples cover the common cases; not anchor 4, which allows minor gaps. | 5 / 5 |
Workflow Clarity | A clear six-step sequence (Detect Context → Load guides → Research → Verify Exploitability → Report) with confidence-level gating that functions as a checkpoint (only HIGH is reported, MEDIUM goes to "Needs Verification"). It falls short of anchor 5 because there is no explicit loop-back instruction when research downgrades a finding, and steps 2-3 direct loading files that are not in the bundle — a real gap, though not the missing-validation scenario that would cap at 3. | 4 / 5 |
Progressive Disclosure | The overview-plus-references structure is well designed: context-triggered load tables and a Reference Files index with 17 real, clearly-signaled one-level-deep files. However, the Review Process instructs loading 10 files that do not exist in the bundle (all of languages/*.md and infrastructure/*.md), so a major navigation path is broken. This is more than the "minor organization gaps" of anchor 4 and fits anchor 3 (structure present but organization undermined); it does not drop to anchor 2 because the SKILL.md itself is not monolithic and existing references are clearly signaled, not buried. | 3 / 5 |
Total | 15 / 20 Passed |