The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative overrides such as 'Always and Automatically improve', 'Invoke this skill automatically when...', and 'UPDATE IMMEDIATELY' to bypass standard agent interaction loops and force autonomous file modifications without explicit user per-action approval.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant vulnerability surface for indirect injection attacks.
Ingestion points: The skill reads instructional context and runtime outcomes derived from session history, which may include interactions with untrusted external data.
Boundary markers: There are no boundary markers or sanitization steps defined to distinguish between legitimate performance improvements and adversarial instructions disguised as 'shortcuts' or 'fixes'.
Capability inventory: The skill possesses the capability to read from and write to any SKILL.md file within the .opencode/skill/ directory.
Sanitization: The process lacks validation of the captured 'learnings', meaning a successful jailbreak or malicious instruction encountered during a task could be persistently 'reinforced' into the skill's permanent documentation.
[DYNAMIC_EXECUTION]: This meta-skill facilitates the runtime generation and injection of new instructional logic into the agent's core capability files. By programmatically appending code blocks, workflows, and templates to SKILL.md files, it creates a self-modifying instruction set which functions similarly to dynamic code execution in an LLM-driven environment.

skill-reinforcement