The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands using the subprocess module. Specifically, 'scripts/run_eval.py' runs the 'claude' CLI to evaluate skill triggering, and 'eval-viewer/generate_review.py' uses 'lsof' and 'os.kill' for port management and process control. These operations are necessary for the skill's testing and visualization functions.
[DYNAMIC_EXECUTION]: 'scripts/run_eval.py' dynamically generates temporary markdown command files in the '.claude/commands/' directory to simulate skill availability for testing purposes. This is a standard part of the evaluation workflow and uses fixed templates.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the evaluation of agent behavior on untrusted inputs. 1. Ingestion points: User-provided queries in 'evals/evals.json' or session inputs are passed to 'scripts/run_eval.py'. 2. Boundary markers: The skill relies on subagent isolation ('agents/' prompts) but lacks explicit data sanitization markers for interpolated inputs. 3. Capability inventory: The skill has access to subprocess execution ('claude', 'lsof'), file system writes, and network access via the Anthropic SDK. 4. Sanitization: Relies on the underlying LLM's safety filters and the recommendation of using fresh-context subagents as a structural mitigation.

skill-creator