The Agent Skills Directory

[SAFE]: The skill defines a cognitive discipline aimed at improving reasoning quality and reducing biases such as anchoring. It does not introduce new tools or capabilities, but rather structures how existing tools are used.- [INDIRECT_PROMPT_INJECTION]: The 'Evidence' phase (Step 3) in SKILL.md ingests data from external files and documentation, which could contain malicious instructions. However, the protocol's mandatory 'Attack' phase (Step 4) requires the agent to generate the strongest possible refutation of its leading hypothesis, providing a structural defense against indirect prompt injection. Mandatory evidence chain: 1. Ingestion points: External files and documentation referenced in Step 3. 2. Boundary markers: None explicitly defined. 3. Capability inventory: Spawning sub-agents via agent() calls and reading/writing scratch files for state management. 4. Sanitization: Not explicitly implemented.- [COMMAND_EXECUTION]: The protocol utilizes a local scratch file to maintain 'Externalized memory', which is a standard context-management practice and does not involve execution of untrusted code or unauthorized data access.- [PROMPT_INJECTION]: The skill instructions include using platform-specific keywords like 'ultrathink' to enhance reasoning effort. These are documented as platform 'dials' for increased reasoning depth and do not constitute an attempt to bypass safety guardrails.

nuke-think