The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and redesign system components based on untrusted project data (code, specifications, and documentation).
Ingestion points: The skill reads repository entrypoints and detail files to build a system model, as described in SKILL.md.
Boundary markers: SKILL.md contains explicit instructions to treat repository text and tool output as evidence rather than as authority to override user instructions or reveal secrets. It also identifies 'poisoned retrieved instructions' as a specific failure scenario to be designed against.
Capability inventory: The agent is authorized to perform documentation edits and, in implement mode, code changes. It also executes a local Python validator for metadata checks.
Sanitization: Instructions in SKILL.md and references/accretion.md emphasize distinguishing observed facts from proposed behavior and ensuring that retained knowledge does not acquire instruction-level authority.
[COMMAND_EXECUTION]: The skill includes a test suite that performs command execution to verify tool functionality.
Evidence: The file tests/test_validate_traceability.py uses subprocess.run to invoke the skill's own scripts/validate_traceability.py script. This execution is confined to verifying the CLI behavior of the included metadata validator during development and testing.

agent-ergonomics