Grants — NIH Funding Intelligence

Portability: Requires bash_tool (for RePORTER POST via curl), Node.js with docx package, and a Consensus MCP connection. Works in Claude Code CLI natively. In Claude.ai with Code Execution + Consensus MCP, the workflow is supported but slower.

Scope: NIH-only. Non-NIH funders (PCORI, DOD CDMRP, VA, foundations) are out of scope and flagged at intake.

For a clinical researcher with a research idea, produce a strategic NIH funding overview as an editable .docx. Output covers research positioning analysis, institute mapping, targeted grant discovery, and strategic recommendations the researcher can edit, copy from, and share with their mentor.

Agent Integrity Rules (Research-Pack Convention)

Inherited; locked verbatim per PR #657 audit.

• Execution discipline. A step isn't complete until result is confirmed received. Consensus calls sequential with 1+ sec pause. RePORTER calls sequential.
• Data sourcing. Count only what tool calls returned this session. Never supplement with training knowledge. Training knowledge labeled [Not from Consensus/RePORTER — reference information] and excluded from counts.
• Counts & attribution. Queries sent / results shown / results cited — three separate numbers, never conflate. Every cited paper has retrievable URL from this session.
• Error handling. On failure → wait 3s → retry once → log. After 3 consecutive failures across tools: stop, alert researcher, explain what's missing. Never silently skip.
• Transparency. Audit Log section in the DOCX. Same standards in chat summary as in document.

See references/reporter_post_patterns.md for the RePORTER POST canon + plan-tier detection.