The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from project/gabbe.config.json and user task intent to determine operational posture.
Ingestion points: Reads configuration parameters and budget limits from project/gabbe.config.json and task descriptions from the user.
Boundary markers: The skill specifies that the configuration must be validated against expected types and values, and uses a structured output format for decisions.
Capability inventory: Accesses the file system (via MCP) to read config and write to AUDIT_LOG.md; coordinates the execution of other skills and model selection.
Sanitization: Explicitly includes a security section stating that hostile or malformed configs must fall back to safe defaults and cannot override safety gates or autonomy levels.
[EXTERNAL_DOWNLOADS]: The configuration example references external registries for skills, including a repository from Google (a well-known service) and https://skills.sh.
The skill mitigates risks associated with these downloads by explicitly requiring human approval for any actions involving externally-sourced code, regardless of the autonomy level.
[SAFE]: The core logic of the skill is defensive, emphasizing 'most-restrictive-wins' for alignment and auditable decision-making. It contains specific instructions to ignore prompt injections that attempt to escalate privileges or bypass constraints.

self-optimize