Content
78%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 impressively lean playbook with concrete sink lists and, notably, an explicit validation section including negative controls — rare for this skill class. The main weakness is that the core payload step is described rather than given: no executable XML payloads (file-read or blind/OOB) and no remediation-check commands, so the reader must construct the actual test payloads themselves.
Suggestions
Add copy-paste-ready payload XML for each case (e.g., a DOCTYPE with a SYSTEM entity for file read, and a parameter-entity payload pointing to an attacker-controlled host for blind/OOB).
Include the concrete checks for "Dangerous defaults" — e.g., the specific Java feature flags (disallow-doctype-decl, external-general-entities) or Python/lxml parser config options to look for.
Add a short error-recovery path when the positive control fails (e.g., retry with parameter entities, OOB DNS/HTTP channels, or try alternate parsers in the same application).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~20-line body is pure signal: named parser classes per language, a checklist of dangerous defaults, one-line payload targets, and a validation section. It assumes Claude's competence (no XML/DTD primer), and every line earns its place — squarely matching the lean-and-efficient anchor. | 5 / 5 |
Actionability | Sink hunting is concrete (specific classes like `DocumentBuilderFactory`, `lxml.etree`), but the payload guidance is descriptive rather than executable — "entity to `file:///etc/passwd`" instead of a copy-paste XML payload with DOCTYPE/ENTITY declarations, and blind/OOB payloads are absent despite the frontmatter advertising them. This lands between "some concrete guidance but incomplete" (3) and "mostly executable" (4); the missing ready-to-use payload XML keeps it at 3. | 3 / 5 |
Workflow Clarity | The section order (find sinks → check dangerous defaults → send payloads → validate) forms a clear, logical sequence, and a Validation section with explicit positive and negative controls is present, satisfying the rubric's requirement for verification in risky operations. It falls short of 5 because steps are implied by headers rather than sequenced, and there is no error-recovery guidance (e.g., what to try when the positive control fails, such as switching to blind/OOB exfiltration). | 4 / 5 |
Progressive Disclosure | The skill is under 50 lines, single-purpose, with no bundle files (references/, scripts/, assets/ are absent) and no external content that needs splitting. Per the rubric's simple-skill guideline, well-organized sections alone qualify for a 5 here. | 5 / 5 |
Total | 17 / 20 Passed |