The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process external data sources to define experimental bounds, which presents a surface for indirect prompt injection.
Ingestion points: In SKILL.md, the agent is directed to "Read the request, project instructions, and existing issue or interview record."
Boundary markers: The skill mitigates risks by requiring the agent to "Establish question, scope/exclusions, scenarios, constraints" and distinguish between "experimental assumptions" and "implemented behavior."
Capability inventory: The workflow involves the agent creating code, modifying files, and interacting with Git and Linear.
Sanitization: The instructions mandate that the agent "Verify execution authorization" and "Isolate the experiment from ordinary use."
[DYNAMIC_EXECUTION]: The primary purpose of the skill is to guide the agent in building and running code experiments, which is a form of dynamic execution.
Evidence: SKILL.md and references/logic.md provide instructions for building variants, running startup scenarios, and verifying experimental behavior.
Mitigation: The skill emphasizes creating the "smallest experiment" in a "representative environment" and using isolation checks to ensure the prototype does not affect production systems.
[COMMAND_EXECUTION]: the skill framework incorporates version control and project management commands.
Evidence: references/retention-and-handoff.md details the use of Git for branch management, commits, and pushing to remotes, as well as updating Linear records.
Security Controls: The skill explicitly warns that "Experiment authorization alone does not authorize a commit" and "never commit secrets, private data, or resources whose publication is not authorized."
[NO_CODE]: The skill consists entirely of Markdown documentation and instructions. It does not contain any scripts, binaries, or automated shell commands that execute upon loading.

prototype