The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from external benchmarks, grader feedback, and session histories to guide improvements.
Ingestion points: Failure signals and feedback sources are ingested into the context as described in SKILL.md and references/evidence-and-failure-classification.md.
Boundary markers: It mandates the use of labels like Primary source and Local investigation to distinguish between external claims and verified facts.
Capability inventory: The skill provides instructions for modifying skill packages and evaluation suites, and invokes a local eval_runner.py utility.
Sanitization: It requires rigorous verification of all external claims against repository artifacts and prohibits trusting relayed summaries without evidence.
[DATA_EXFILTRATION]: The skill establishes clear boundaries for sensitive data handling.
It explicitly forbids the transmission of private skill packages or session excerpts to external services without explicit authorization in references/eval-quality-and-proof.md.
[COMMAND_EXECUTION]: The skill references a tool for running evaluations (eval_runner.py).
It includes safety constraints that prevent unauthorized execution, requiring explicit user triggers for all evaluation runs.

skill-quality