The Agent Skills Directory

[SAFE]: The skill establishes a disciplined research methodology for AI agents, prioritizing data integrity and experiment validation. A thorough audit of the instructions and reference files revealed no malicious intent, unauthorized network exfiltration, or obfuscation techniques.
[COMMAND_EXECUTION]: The skill instructions and verification gates (G1-G8) make use of common local command-line utilities such as grep, jq, diff, and git to compare experiment manifests and search for project claims. These operations are performed on local files for the purpose of validating research consistency.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process research data, experiment logs, and user-provided configurations, which represents a potential injection surface. However, the skill explicitly mandates robust safety and integrity measures to mitigate this risk.
Ingestion points: Research logs (records/ledger.tsv), project claims (records/CLAIMS.md), and experiment manifests (manifest.json).
Boundary markers: Implements a strict '3-layer separation' model (Ledger, Claims, and Constitution) to isolate raw data from high-level project principles.
Capability inventory: Local file searching and structured data comparison using grep, jq, and diff.
Sanitization: Employs eight mandatory launch gates (G1-G8) to verify that data inputs, experiment deployments, and results align with the human-approved 'INTENT.md' and project constitution.
[EXTERNAL_DOWNLOADS]: The skill contains references to external scientific resources from well-known services, including Cornell University's arXiv repository and METR's research blog. These references are used for academic context and do not involve the execution of untrusted code.

stacked-research