The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external documents, metrics, and plans provided by the user. This ingestion surface is a potential vector for indirect prompt injection if the inputs contain malicious instructions. However, the skill implements multiple security controls to mitigate this risk: it requires a formal 'Input Audit' to categorize data (known, missing, conflicting, stale), mandates the separation of facts from inferences, and lacks active capabilities such as network access or file writing. These constraints significantly limit the impact of any potential injection.
Ingestion points: prompts/quality-productivity-metrics.md (Input Audit section) and SKILL.md (How to Use section).
Boundary markers: The prompt explicitly directs the agent to separate facts from inferences and recommendations, and to identify 'out_of_scope' content.
Capability inventory: The skill logic is confined to analysis; no network operations or file system modifications are present in the core instructions.
Sanitization: Instructions include using 'least privilege, masked data, mocks, dry runs, or isolation' for sensitive work.
[COMMAND_EXECUTION]: The evaluation framework includes a shell script and a Python script used to verify the skill's logic during development. These scripts are executed locally within the evaluation context and do not process untrusted external code.
Evidence: evals/fixtures/check-domain-decisions.sh executes a local Python check.

quality-productivity-metrics