The Agent Skills Directory

[PERSISTENCE]: The skill implements a 'Distill' procedure that proposes updates to the agent's skill library on disk. This mechanism is designed for building long-term memory but is strictly gated by a requirement for explicit user confirmation before any modifications are committed.
[DATA_EXFILTRATION]: The instructions describe an optional image review process that uploads renders to a server specified by '$IMAGE_SERVER_URL'. The skill enforces a strict safety rail that the URL must be localhost-bound (defaulting to http://localhost:8092) and requires user approval for each upload.
[COMMAND_EXECUTION]: The framework discusses patterns for fleet management and deployment using SSH and shell scripts. These are presented as high-level architectural guidelines for managing robotics simulation environments rather than direct, hidden command execution.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves researching and processing external documentation to generate new instructions.
Ingestion points: Research resources, user requests, and task outputs are ingested during the Discovery and Practice phases (SKILL.md).
Boundary markers: Mandatory user confirmation serves as a check before any generated content is saved.
Capability inventory: The skill is capable of proposing file writes and network uploads to local servers.
Sanitization: The framework mandates a 'Validation' phase and 'Quality Gates' with checklists in references/quality-criteria.md to verify generated content against established rubrics.

meta-skills