The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data, including legacy code, target instructions, and external evidence, to establish behavioral baselines.
Ingestion points: The skill processes external inputs in SKILL.md (Step 1: target instructions, affected code; Step 2: representative inputs).
Boundary markers: The instructions emphasize establishing "boundary and authority" and using explicit dispositions (Intended, Unknown, Defect) to categorize observations.
Capability inventory: The agent is authorized to "Add the smallest behavioral harness" and "Run existing relevant checks" (SKILL.md Step 3), and can perform Git writes via specific companion tools like implementation-preflight or work-on. This creates a multi-step chain where output from one step influences the next.
Sanitization: The skill lacks explicit instructions for sanitizing or escaping instructions found within the legacy code or documentation it processes.
[DYNAMIC_EXECUTION]: The skill generates and executes code at runtime to perform characterization tests.
Evidence: SKILL.md Step 3 instructs the agent to create a "behavioral harness at the real boundary" to capture baseline results.
Evidence: The provided scripts/check-fixture.py demonstrates this capability by dynamically writing a synthetic legacy script to a temporary file for execution.
[COMMAND_EXECUTION]: The skill utilizes system commands to execute the generated test scripts.
Evidence: scripts/check-fixture.py uses the subprocess.run function to invoke the Python interpreter and run diagnostic checks against temporary files.
[EXTERNAL_DOWNLOADS]: The skill's documentation provides installation instructions that involve downloading the skill package from an external repository associated with the author.
Evidence: The README.md suggests using npx skills add SirDarcanos/agentskills to install the skill and its dependencies.

characterize