The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided content, including evaluation prompts, test datasets, and feedback files. This data is subsequently passed to secondary agents (graders and analyzers), creating a surface where instructions embedded in the processed data could potentially influence the subagents' behavior.
Ingestion points: Processes evals/evals.json, feedback.json, and output files generated during test runs in scripts/run_loop.py and eval-viewer/generate_review.py.
Boundary markers: The skill instructions emphasize the use of distinct sections in SKILL.md and recommend the use of table of contents for large reference files to maintain context organization.
Capability inventory: The skill environment has the capability to execute shell commands (via subprocess), perform file system operations (read/write), and host a local HTTP server.
Sanitization: Data is structured into JSON for the results viewer and passed through standard CLI argument patterns.
[COMMAND_EXECUTION]: The toolkit uses Python scripts to automate interaction with various developer CLI tools. Scripts like run_eval.py, improve_description.py, and run_loop.py utilize the subprocess module to execute commands for the claude, codex, opencode, and agy harnesses.
Evidence: Use of subprocess.Popen and subprocess.run to invoke external binary harnesses for benchmarking and description optimization.
[EXTERNAL_DOWNLOADS]: The results viewer component fetches a standard JavaScript library from a well-known content delivery network to provide spreadsheet rendering capabilities.
Evidence: eval-viewer/viewer.html includes a script tag loading the SheetJS library from cdn.sheetjs.com with a Subresource Integrity (SRI) hash.

skill-creator