The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites via WebFetch, which is a standard surface for indirect prompt injection in research-oriented skills.
Ingestion points: Data enters the agent's context during Phase 3, where WebFetch is used to retrieve content from search results.
Boundary markers: The skill uses a structured extraction prompt (prompts/page-extract.md) to separate content into categories (Key Findings, Code Examples, etc.), which helps organize the data but does not fully isolate the model from malicious instructions embedded in the source text.
Capability inventory: The skill can perform network searches and write documentation files to the docs/research/ directory.
Sanitization: The skill's security section explicitly mentions validating URLs and sanitizing sensitive paths before saving documentation.
[SAFE]: The skill implements robust security controls and boundaries to prevent unauthorized actions.
Veto Conditions: It features specific triggers to block implementation requests (code generation, deployment) and attempts to write files to unauthorized paths (e.g., .claude/agents/, src/).
Scope Boundaries: Execution is strictly limited to research and documentation, with no provision for executing bash commands or installing external dependencies.

tech-search