The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing local repository scripts such as scripts/probe-skill.sh and scripts/check-skill-probe-headroom.sh (referenced in SKILL.md and references/behavioral-probes.md), along with the claude plugin eval command. These executions are integral to its function as a testing framework and operate within the expected environment of a development repository.\n- [INDIRECT_PROMPT_INJECTION]: The skill evaluates outputs from other AI agent skills, which constitutes an ingestion point for potentially untrusted data. The instructions mitigate this by requiring fixed evaluation criteria, discriminator calibration, and paired-outcome analysis. The attack surface is inherent to the purpose of an evaluation tool.\n
Ingestion points: Evaluated skill outputs via claude plugin eval and scripts/probe-skill.sh (as described in SKILL.md Procedure step 5).\n
Boundary markers: Isolation is managed by freezing task definitions and environment variables across comparison arms (SKILL.md Procedure step 3).\n
Capability inventory: Shell execution of scripts in scripts/ and the claude CLI; file writing for probe-result.v1 and internal reports (SKILL.md Frontmatter and Runners section).\n
Sanitization: Requires the use of sanitized fixtures for any data committed to the repository (SKILL.md Procedure step 7).\n- [EXTERNAL_DOWNLOADS]: The skill documentation notes that the claude plugin eval tool publishes reports to claude.ai by default. This refers to a well-known service associated with the agent platform and is documented as standard behavior.

skill-eval