grants – skill Dữ liệu & Khoa học cho agent AI

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](references/reporter_post_patterns.md) for the RePORTER POST canon + plan-tier detection.

Phase 1: Grill-Me Intake (6 forcing questions, one at a time)

Q1 (root) — Research idea

> Describe the research idea in 2–3 sentences. What's the question, what's new, and what's the clinical relevance? Vague answers ("AI for healthcare", "biomarkers for disease X") will be rejected — push for specificity.

>

> *Why I'm asking:* Five Consensus searches (established / stakes / current approaches / adjacent methods / gaps) depend on a precise research idea. Vague ideas produce vague gap quotes and useless positioning narrative.

Refuse mush. Re-ask once with examples if user is too broad.

Q2 (depends on Q1) — Career stage

> Career stage — pick one:

>

> 1. Pre-doctoral (PhD student, T32 trainee)

> 2. Postdoctoral fellow (F32, K99 candidate)

> 3. Early career (K-award candidate, first R01)

> 4. Independent investigator (multiple R01s, established lab)

> 5. Senior PI (R35, P-series, U01 leadership)

>

> *Why I'm asking:* Career stage filters mechanism recommendations. F-series for trainees, K-series for early career, R-series for independent. Picking the wrong stage produces unfundable mechanism suggestions.

Forcing choice.

Q3 (depends on Q2) — Preliminary data status

> Preliminary data — pick one:

>

> 1. None (de novo project, no pilot data yet)

> 2. Pilot data (early findings, single-site)

> 3. Strong preliminary (multi-experiment, ready for R01-scale)

> 4. Validated and ready (multi-site, publication-ready)

>

> *Why I'm asking:* Prelim data status drives mechanism budget. No data → R03 / R21 pilot scope. Strong prelim → R01 / U01 multi-site scale. Mismatch produces uncompetitive applications.

Q4 (depends on Q2) — Environment

> Research environment — pick one:

>

> 1. R01-eligible (research-intensive institution with NIH base funding)

> 2. Mid-tier (regional academic medical center, modest NIH portfolio)

> 3. Resource-constrained (smaller institution, minimal NIH base)

> 4. Industry-collaborative (academic + industry partnership)

>

> *Why I'm asking:* Environment affects scope realism (multi-site U01 requires R01-eligible) and which mechanism categories are competitive (R15 specifically targets resource-constrained).

Q5 (depends on Q1) — Submission posture

> Submission posture — pick one:

>

> 1. New application (first submission, no prior reviews)

> 2. Resubmission (A1 with reviewer responses needed)

> 3. Exploring (haven't decided yet whether to submit)

>

> *Why I'm asking:* Resubmissions need reviewer-response guidance in the DOCX (Section 7). New applications skip that. Exploring shifts emphasis to landscape over strategy.

Q6 (depends on Q1) — Known institute targets

> Are you already considering specific NIH institutes? List names (NCI / NHLBI / NIMH / NINDS / NIDDK / etc.) or say "no preference — find the right ones".

>

> *Why I'm asking:* If you have an institute hypothesis, I'll validate it against RePORTER data. If not, I'll surface the top-3 institutes funding adjacent work from the institute-tally.

Accept "no preference" as the common case.

Stop condition: After Q6, commit and start Phase 2A. Never re-open intake after Phase 2A begins.

Phase 2A: Research Positioning (5 Consensus searches)

Run sequentially at 1 q/sec. Each search corresponds to one positioning facet:

1. Established — "<research idea>" established evidence — what's known

2. Stakes — "<topic>" mortality OR burden OR cost OR prevalence — why it matters

3. Current Approaches — "<topic>" current treatment OR standard of care OR approach — state of the art

4. Adjacent Methods — "<related technique>" applied to <topic> — methodological possibilities

5. Gaps — "<topic>" limitations OR unanswered OR future directions OR challenge — gap signals

Use scripts/citation_tracker.py --action record_consensus_search for each. Plan-tier detected from first response.

Synthesis: for each facet, extract 2-3 quotable findings (becomes Section 2 gap quotes). Draft Significance/Innovation language using "the field has established X (refs), but Y remains unanswered (refs)" pattern.

Phase 2B: Institute Mapping + Grant Discovery (RePORTER POST)

RePORTER is POST-only. Use bash_tool + curl — never web_fetch.

Dynamic fiscal year window

Compute at runtime via scripts/fiscal_year_calculator.py. Default: current FY + 3 prior. Federal FY starts Oct 1, so:

python scripts/fiscal_year_calculator.py --output json
# Returns: {"current_fy": 2026, "window": [2023, 2024, 2025, 2026]}

Narrow (AND) search — finds direct overlap

curl -X POST 'https://api.reporter.nih.gov/v2/projects/search' \
  -H 'Content-Type: application/json' \
  -d '{
    "criteria": {
      "fiscal_years": [2023, 2024, 2025, 2026],
      "include_active_projects": true,
      "advanced_text_search": {
        "operator": "AND",
        "search_field": "all",
        "search_text": "<key term 1> <key term 2>"
      }
    },
    "limit": 50,
    "include_fields": ["project_num", "project_title", "agency_ic_admin", "study_section", "fiscal_year", "principal_investigators", "abstract_text"]
  }'

Broad (OR) search — finds adjacent work

curl -X POST 'https://api.reporter.nih.gov/v2/projects/search' \
  -H 'Content-Type: application/json' \
  -d '{
    "criteria": {
      "fiscal_years": [2023, 2024, 2025, 2026],
      "advanced_text_search": {
        "operator": "OR",
        "search_field": "all",
        "search_text": "<term> <synonym> <related concept>"
      }
    },
    "limit": 50
  }'

Institute tally + study section ranking

After RePORTER responses:

• Tally agency_ic_admin (institute code: NCI, NHLBI, NIMH, etc.) → top-3 funding institutes
• Tally study_section → top-2 study sections (where applications go for review)

NOSI discovery

Parse RePORTER responses for NOT-* opportunity numbers. For each:

# NOSIs live at predictable URLs:
# https://grants.nih.gov/grants/guide/notice-files/NOT-<INSTITUTE>-<YEAR>-<NUMBER>.html
web_fetch <url>

If fetch fails: log [NOSI {number} — fetch failed, not included], continue.

Mechanism Matching (Scope-Aware)

NOT career stage alone. Career stage + project scope + prelim data drive recommendation.

Use scripts/mechanism_matcher.py:

python scripts/mechanism_matcher.py \
  --career-stage "early_career" \
  --prelim-data "pilot" \
  --environment "r01_eligible" \
  --scope "single_site" \
  --output json
# Returns mechanism shortlist with rationale

See [references/nih_mechanism_matching.md](references/nih_mechanism_matching.md) for the full matrix.

Phase 3: DOCX Generation

9 sections via Node.js + docx library. See [references/docx_9_sections.md](references/docx_9_sections.md) for full spec.

1. Executive Summary — title + career stage + environment + 3-4 key findings bullets

2. Research Positioning — 3-5 gap quotes (italicized, inline Consensus citations) + 2-3 paragraph positioning narrative + supporting evidence table

3. Target Institutes — ranking table (institute, project count in window, % match to your idea) + 2-3 sentence interpretation

4. Grant Opportunities — bold NOSI callout if any

… (nội dung đầy đủ có trong gói tải về)