The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use local Python scripts (scripts/skillctl.py and scripts/install-external-skills.py) for operational tasks. These include classifying differences in code, proposing upstream patches, and managing external skill installations.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of external AI agent skills (such as writing-great-skills or teach) through a local registry and installation script. This is standard functionality for skill management within this ecosystem.
[CREDENTIALS_UNSAFE]: The skill metadata (.skillmeta.yml) and instructions include explicit protections against credential exposure. Files such as personalize.local.yml, *.local.yml, and directories like .company/ or .user/ are marked as forbidden for upstream contributions to ensure private user and environment data remain local.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a mutation and evaluation engine for other prompts, which inherently creates an attack surface for processing potentially malicious instructions.
Ingestion points: Processes skill diffs and mutation proposals via scripts/skillctl.py and evaluates skill behavior using user-provided eval sets.
Boundary markers: The documentation in references/skill-eval-methods.md mandates strict human-in-the-loop approval gates, regression checks, and evidence-based shipping criteria to mitigate the risk of unintended behavior changes.
Capability inventory: Execution is scoped to Python-based helper scripts for diffing, patching, and installation.
Sanitization: Relies on human oversight and explicit regression gates rather than automated sanitization, which is appropriate for a development-focused evaluation loop.

skill-eval-loop