The Agent Skills Directory

[PROMPT_INJECTION]: Explicit instructions command the agent to bypass user interaction and avoid pauses or confirmation requests during the autonomous evaluation loop.
Evidence: SKILL.md Phase 4b and references/eval-loop.md specify 'do not pause to ask the user' and 'DO NOT ask for confirmation'.\n- [COMMAND_EXECUTION]: The skill executes local system commands and scripts to manage the environment and score results.
Evidence: Runs 'scripts/score_evals.py' via the Python interpreter and performs multiple 'git' operations including 'init', 'add', 'commit', 'reset', and 'checkout'.\n- [REMOTE_CODE_EXECUTION]: The skill dynamically generates or modifies agent instruction files and executes them via the Agent tool to verify behavior.
Evidence: Rewrites 'SKILL.md' or prompt files and uses the 'Agent' tool to execute them against test prompts in Phase 4b.\n- [EXTERNAL_DOWNLOADS]: Fetches data from external web sources to gather best practices for rewriting artifacts.
Evidence: Uses 'WebSearch' and 'WebFetch' to retrieve information from Anthropic's official documentation and engineering blogs.\n- [COMMAND_EXECUTION]: Modifies file system permissions and attributes to protect configuration files during the evaluation process.
Evidence: Executes 'chmod 444' on Unix and 'attrib +R' on Windows for 'evals.json'.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the ingestion of untrusted external data and local artifact content.
Ingestion points: Reads local artifact files (SKILL.md, CLAUDE.md) and external web content from 'WebFetch' results.
Boundary markers: No explicit delimiter or 'ignore instructions' markers are used for processed data.
Capability inventory: Performs file system writes, git command execution, and recursive agent tool calls.
Sanitization: No sanitization or validation of the ingested content is performed before interpolation into prompts.

auto-research