Content
77%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 highly actionable, well-sequenced chaining playbook whose main weaknesses are redundancy across the pattern/conditional-table/tier/quick-reference sections and a monolithic single-file structure with no progressive disclosure. The workflow, decision checkpoints, and concrete commands are excellent.
Suggestions
Deduplicate the open redirect/takeover/SSRF chains that currently appear in both the 15 patterns and the Conditional Chain Table, Payout Tiers, and Quick Reference — keep one canonical pattern description and have the other sections reference it by pattern letter only.
Split the pattern catalog (A–O), per-class escalation ladders, and CVSS multiplier table into references/ files (e.g. references/patterns.md, references/escalation-ladders.md), keeping the filter, workflow, clock, and conditional table in SKILL.md as the overview.
Trim the per-pattern 'Payout range' annotations in the pattern bodies since the same figures are consolidated in the PAYOUT TIERS section.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and mostly free of concept explanations Claude already knows, but the same chain material repeats across sections — e.g. open redirect → OAuth code theft appears in Pattern C ("Craft: /oauth/authorize?redirect_uri=https://target.com/redirect?to=https://evil.com"), the Conditional Chain Table ("Open redirect | + OAuth redirect_uri → code theft"), Payout Tier B ("Open redirect → OAuth code theft → ATO"), and the Quick Reference ("OAuth callback | Open redirect on any page → code theft (Pattern C)") — and payout figures are stated twice (per-pattern and in tiers). Consolidating these would cut meaningful tokens without losing information. | 3 / 5 |
Actionability | Guidance is copy-paste ready throughout: exact parameters ("redirect=, return_to=, next=, url=, r=, continue="), exact metadata endpoints ("http://169.254.169.254/latest/meta-data/iam/security-credentials/" with IMDSv2/GCP/Azure variants), concrete payloads ("PATCH /api/users/{victim_id} {\"role\":\"admin\"}"), specific tools and flags ("hashcat -a 0 -m 16500", "Turbo Intruder (single-packet)"), and named mass-assignment parameters to test. For an instruction-only skill this is as actionable as it gets. | 5 / 5 |
Workflow Clarity | The CHAIN-FINDING WORKFLOW gives a clear classify → find gadget → prove → maximize sequence with explicit go/no-go checkpoints ("If you cannot write Step 2 as a real HTTP request → the chain isn't proven. Do not submit."), the 10-question filter is a scored checklist with decision rules ("3+ YES → ready to report... 0 YES → kill it"), and the 30-minute clock adds time-boxed decision points with feedback loops ("If stuck: 10-question filter — which Q can you make YES?"). | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent), so everything lives in one ~500-line SKILL.md. Internal section structure and headers are good, but content that naturally belongs in reference files — the 15 pattern write-ups (~200 lines), the escalation ladders, and the CVSS multiplier table — is fully inlined rather than split out, leaving SKILL.md as a catalog rather than an overview. | 3 / 5 |
Total | 16 / 20 Passed |