The same hardening code recurs roughly three times (Patterns That Prevent XXE, Vulnerable vs. Secure Examples, and the Execution recon list — e.g. Java DOM, lxml, .NET, and PHP each repeated), and the 'What is XXE' section explains a concept Claude largely already knows; not the verbose-walled level 1, but not the lean level 3 either.

Provides executable per-language code with exact feature flags and copy-paste-ready curl payloads in the output templates, fully matching the 'executable, copy-paste ready' anchor.

Three phases are explicitly sequenced with an between-phase validation gate (zero-findings early exit writes 'No vulnerabilities found.'), parallel batching rules, merge-after-all-complete ordering, and an 'Important Reminders' checklist; the batch operation includes validation so the level-2 cap does not apply.

No bundle files exist and the ~550-line body is monolithic, with large language-by-language pattern catalogs that could be split into reference files; sections are well-organized (not level 1) but content that should be separate is inline (not level 3), and the under-50-line exception does not apply.

Lists multiple concrete actions across a named three-phase approach: 'recon (find XML parsing sites without external-entity hardening), batched verify (trace user input to each site in parallel subagents, 3 sites each), and merge (consolidate batch results)', which exceeds the 'names domain and some actions' level.

It explicitly answers 'what' (detect XXE via recon/verify/merge with named outputs and prerequisite) and 'when' via an explicit 'Use when...' clause, matching the top anchor; not capped at 2 since the trigger clause is present.

The clause 'Use when asked to find XXE or XML injection bugs' supplies natural phrasings a user would actually say, avoiding generic jargon-only triggers.

Scoped narrowly to XXE/XML-injection detection with distinct triggers and an explicit sibling-skill prerequisite ('Requires sast/architecture.md'), making overlap with other skills unlikely.