The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill follows security best practices for credential management by automatically identifying sensitive variable names (such as tokens, passwords, pins, and keys) and marking them as secrets with empty values in the generated environment configuration files.
[REMOTE_CODE_EXECUTION]: Code execution is limited to local Python scripts (colcheck.py and yaml2md.py) included as assets within the skill. No remote downloads or untrusted script executions were found.
[DATA_EXFILTRATION]: The skill operates entirely on local project files within the docs/api, docs/design, and collection directories. No network activity or external data transmission was identified.
[PROMPT_INJECTION]: Instructions for both the main agent and the independent verification agent are structured and task-oriented, with no evidence of bypass markers or attempts to override AI safety guidelines.
[SAFE]: The skill manages potential indirect prompt injection surfaces through the following architecture:
Ingestion points: Data is ingested from local docs/api/*.yaml and docs/design/*.html files.
Boundary markers: The generator uses rigid, pre-defined templates to isolate document content within specific YAML scalar blocks.
Capability inventory: The skill uses local scripts for mechanical verification and a subagent for semantic verification, with all actions confined to the project workspace.
Sanitization: The yaml2md.py asset ensures that data from the specifications is properly wrapped in markdown code blocks or escaped before inclusion in the final output.
[SAFE]: Data integrity is maintained via a three-layer verification process, including a deterministic script tripwire, an independent 'fresh-eyes' agent review, and a final completeness sweep to catch silent omissions.

open-collection