The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection as it is designed to ingest and process target agent skills, which are potentially untrusted third-party data.\n- Ingestion points: Target skill directories, specifically SKILL.md, references/, and scripts/ are read during Phase 0 (Target Acquisition).\n- Boundary markers: The skill explicitly defines 'Hard Rules' and 'Gate Checkpoints' that the agent must follow, providing structural boundaries against adversarial instructions.\n- Capability inventory: The agent is granted file read/write access to local files and has the capability to execute scripts during the validation phases.\n- Sanitization: All modifications require an explicit 'User approval required' gate (Gate 3). Additionally, the skill utilizes 'Self-Correction Rules' (Phase 4) and a final 'Spec Validation' check (Phase 5) to ensure resulting output conforms to safe and expected formats.\n- [DYNAMIC_EXECUTION]: The skill workflow includes a verification phase where the improved skill and a baseline version are executed to compare behavior.\n- Evidence: references/trial-run.md provides instructions for running both baseline and improved versions of the target skill's scripts within a fresh context to gather objective quality evidence.\n- Context: This dynamic execution is a documented feature of the skill's primary audit function and is applied solely to the local skill files identified by the user.

improve-skill