The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a 'Skill generation loop' where the agent is instructed to execute generated code solutions to verify correctness and run local validation scripts (./ci/utils/validate_skills.sh and ./ci/test_skills_assets.sh) before proposing updates.
[PROMPT_INJECTION]: The skill facilitates a mechanism for indirect prompt injection (Category 8) by modifying its own instructions based on external data from user interactions. \n
Ingestion points: User corrections and problem-solving retries are ingested as training data for skill updates. \n
Boundary markers: The skill uses explicit boundary markers () to tag and delimit generated content for review, though it lacks specialized delimiters for incoming user feedback. \n
Capability inventory: The skill possesses the capability to modify local instruction files (SKILL.md), write new Python scripts (assets/*.py), and execute shell-based validation tools. \n
Sanitization: Security is maintained through a mandatory adversarial reasoning checklist, a requirement for explicit user approval before any changes are persisted, and the enforcement of non-negotiable rules that prevent the adoption of instructions which would expand permissions or weaken safety filters.

skill-evolution