The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill framework is designed to ingest and act upon data collected during agent operation, such as execution logs, manual corrections, and user overrides.
Ingestion points: External data enters the agent context from execution feedback, stored logs (like mistakes.md and patterns.md), and memory-based metrics.
Boundary markers: The instructions lack explicit delimiters or warnings to treat ingested feedback as untrusted, which could allow malicious instructions in processed data to influence the evolution process.
Capability inventory: The skill enables the agent to perform file-write operations to update SKILL.md files across the environment.
Sanitization: There is no documentation of validation, filtering, or escaping of feedback data before it is incorporated into proposed skill modifications.
[DYNAMIC_EXECUTION]: The skill describes a process for self-modifying instructions where the agent generates and applies code-like updates to its own environment.
Evidence: The evolution process includes steps to "PROPOSE: Generate minimal diff (FIX), new variant (DERIVED)" and "APPLY: Update SKILL.md". This constitutes dynamic generation and application of agent instructions based on runtime data.

skill-evolution