Content
70%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 deeply actionable, well-gated exploit workflow whose confirm-desync gate and validation discipline are exemplary. The main costs are token efficiency (the same anti-brute-force and timeout rules are repeated multiple times) and the total absence of progressive disclosure — a 650-line monolith where three or four chapters belong in separate reference files.
Suggestions
State the 'smuggling beats credential brute-force when tags co-occur' rule once (in When This Skill Is Primary) and reference it elsewhere in one line instead of restating the full reasoning in the Class A anti-pattern and 'Why this beats brute-force' sections.
Split the Class C chapter, the triple-tier desync matrix, and the PortSwigger lab pattern reference into references/ files (e.g. references/class-c-diagnostics.md, references/triple-tier.md, references/labs.md), keeping SKILL.md as a gate + variant-catalog overview.
Complete the fragmentary snippets (TE.CL, B1, B2) with the same send/receive loop used in the gate, or fold them into a single reusable harness function so every payload block is copy-paste executable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — the harnesses, gates, and catalogs are operational content Claude does not already know — but noticeably padded in places: the 'brute-force is a fallback only after the confirm-desync gate fires NEGATIVE' rule is restated near-verbatim in three sections (When This Skill Is Primary, the Class A ANTI-PATTERN block, Why this beats brute-force), and timeout discipline is repeated across Variant Catalog, Anti-Patterns, and Class C. Fits the 'mostly efficient but could be tightened' anchor, not the verbose anchor since the majority of tokens are non-generic operational detail. | 3 / 5 |
Actionability | The confirm-desync gate, Class A detection, and cookie-replay blocks are fully executable copy-paste Python/bash with timeouts and bounded recv loops, and the TE.TE obfuscation table gives exact header lines. Minor gaps: the TE.CL, B1, and B2 snippets are payload-construction fragments without send/receive code ('# Send + read 2 responses'), and B3 plus the sibling-resource enumeration loop are comment-only skeletons. 'Mostly executable guidance with minor gaps' — not 5 because several blocks require the reader to reassemble the surrounding harness. | 4 / 5 |
Workflow Clarity | Excellent sequencing with explicit validation checkpoints and feedback loops: recognition signals → a mandatory confirm-desync gate with enumerated pass criteria and a fail-pivot ('do NOT iterate variants... Hand back to recon') → variant catalog → a Class A/B routing decision tree with run-FIRST detection probes → a verification section requiring 'Repro the smuggle three times' and explicit acceptable evidence. Stop rules and time-boxes ('each is a 5-minute time-box', 'if 8+ outer-pair variants... STOP') make error recovery unambiguous. | 5 / 5 |
Progressive Disclosure | Section headers and navigation within the file are good, but there are no bundle files at all: ~650 lines are entirely inlined, and the Class C chapter, the triple-tier desync matrix, and the PortSwigger lab pattern reference are content that clearly belongs in references/ files. Fits 'some structure but content that should be separate is inline'; not 2 because the file is well-organized with clear headers rather than a headerless wall of text. | 3 / 5 |
Total | 15 / 20 Passed |