The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted repository files and knowledge concept definitions.
Ingestion points: Processes files within the .lokf/knowledge/ directory and arbitrary repository files referenced by a concept's resource field (e.g., SKILL.md).
Boundary markers: Includes specific directives to treat all fetched content as text for summarization and to ignore any instructions embedded within source files.
Capability inventory: Reads repository files and writes feedback logs to .lokf/feedback.md.
Sanitization: Explicitly forbids the inclusion of secrets, credentials, or tokens in outputs, requiring the agent to only describe the location and type of any sensitive data found.
[COMMAND_EXECUTION]: Uses CLI tools to identify the local user for feedback attribution.
Evidence: Executes gh api user --jq .login or glab api user to retrieve the current session's username. This is a static command used for identification and does not involve user-controlled input interpolation.
[EXTERNAL_DOWNLOADS]: References external documentation and plugins maintained by the skill author.
Evidence: Links to the noelmcloughlin/knowledge-trust-ladder repository for examples and noelmcloughlin/obsidian-lokf-registrar for Obsidian integration instructions. These are neutral references to the project ecosystem.

lokf-docent