The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources including the wiki/ directory, queue.md, and existing initiative files. This represents an attack surface where malicious instructions embedded in these files could influence agent behavior.
Ingestion points: Reads project documentation and history from wiki/, tasks from queue.md, and metadata from wiki/initiatives/ files.
Boundary markers: The skill explicitly operates under a "shared contract" and utilizes a "Universal preflight" check to establish a safe operating context.
Capability inventory: The skill has the ability to write multiple project files (wiki/initiatives/, log.md, index.md, hub/initiatives.md, baton.md) and execute a local audit script (scripts/graph-audit).
Sanitization: It implements a "shared durable claim admission gate" to verify facts and a "shared parameter gate" to validate numeric and temporal constraints in the ingested data.
[DYNAMIC_EXECUTION]: The skill requires the execution of a local project script, scripts/graph-audit, as a mandatory final step for every file modification.
Evidence: The instructions in Step 5 mandate running scripts/graph-audit on every touched durable file and require the script to exit with a zero status before the task is considered complete.

plan