The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from local markdown files to generate and execute file modification plans. Because these local files could be influenced by untrusted external data or other agent actions, this represents a potential surface for indirect prompt injection. 1. Ingestion points: The skill reads problem descriptions and optimization suggestions from the docs/optimization/pending-issues.md file as described in SKILL.md. 2. Boundary markers: The system implements a mandatory user confirmation mechanism (y/n/s/a) which requires an explicit human decision before any modifications are applied. 3. Capability inventory: The skill has the capability to read from and write to any project file, including other skill definitions (e.g., .claude/skills/proj-gen/SKILL.md). 4. Sanitization: The skill does not perform automated sanitization or filtering of the input data from the issues log, relying instead on the user's review of the optimization plan preview.
[DYNAMIC_EXECUTION]: The skill generates optimization schemes that involve rewriting or modifying other skill files and documentation. When these modified files are subsequently loaded or used by the agent, it constitutes the execution of dynamically generated or altered instructions.

proj-optimize