The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface via codebase content and user instructions.
Ingestion points: Processes git diff output via scripts/diff-scan.sh and user-supplied instructions from the -m argument in SKILL.md.
Boundary markers: The prompt template in references/test-plan.md uses structural section headers to delimit data, which helps provide context but does not fully prevent adversarial influence from processed data.
Capability inventory: The skill possesses the ability to execute shell commands through the Bash tool and perform browser actions (including JavaScript execution) via agent-browser.
Sanitization: Untrusted data from the codebase and user input is interpolated into the AI test plan generation prompt without explicit sanitization or escaping.
[EXTERNAL_DOWNLOADS]: Fetches configuration and tools from well-known and trusted sources.
Recording Library: The skill optionally downloads the rrweb session recording library from the jsDelivr CDN (cdn.jsdelivr.net) as described in references/rrweb-recording.md.
Automation Tool: The skill documentation in SKILL.md identifies a dependency on the @anthropic-ai/agent-browser package for browser orchestration.
[COMMAND_EXECUTION]: Orchestrates a multi-phase testing pipeline using local scripts and standard system tools.
Pipeline Scripts: Executes several bundled bash scripts (scripts/diff-scan.sh, scripts/fingerprint.sh, scripts/init.sh, scripts/route-map.sh, scripts/report.sh) to manage the testing lifecycle.
Dynamic Logic: Shell scripts utilize inline Python commands (python3 -c) to handle JSON/YAML parsing and framework detection logic.

expect