The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes 'improvement suggestions' fetched from a Supabase database to modify the agent's core configuration and behavior. This pipeline could be exploited if malicious instructions are introduced into the learned patterns via previous interactions with untrusted data.
Ingestion points: Data retrieved via SQL from the v_improvement_suggestions view in SKILL.md.
Boundary markers: The skill requires user approval for changes and performs impact assessments, but does not use explicit delimiters or 'ignore instructions' markers when presenting suggested improvements to the agent for processing.
Capability inventory: File writing to ccpm-config.yaml and agents/router.md, and SQL execution (SELECT, UPDATE) in SKILL.md.
Sanitization: The skill performs syntax validation of generated rules and conflict detection before application.
[COMMAND_EXECUTION]: The skill instructions define specific shell-like triggers (/learn:apply, /learn:rollback) and SQL operations to manipulate configuration files and database records as part of its operational lifecycle.

self-improve