The Agent Skills Directory

[SAFE]: The skill instructions mandate a read-only posture and require explicit user authorization for persistence or skill creation. Hard rules explicitly prevent silent modifications to the repository or creation of skills.
[INDIRECT_PROMPT_INJECTION]: 1. Ingestion points: The skill reads and analyzes repository files and session history ($seq). 2. Boundary markers: The skill establishes explicit categories for evidence (e.g., observed fact vs inference) and includes warnings against 'contamination' from prior agent-generated reports. 3. Capability inventory: The skill is strictly read-only for repository assets and only recommends rather than implements changes. 4. Sanitization: Findings must be re-falsified against current code and executable proof to prevent manipulation.
[COMMAND_EXECUTION]: The skill interacts with platform-provided evidence tools like $seq and $negative-ledger. These interactions are scoped to read-only discovery tasks within the induction workflow, preventing the execution of arbitrary or malicious shell commands.

codebase-doctrine