AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

grants

Use for any NIH funding question ('find grants for my research', 'which NIH institute'). Maps the idea to institutes, study sections and NOSIs via RePORTER, matches mechanisms with mechanism_matcher.py and writes a .docx strategy. NIH only.

Original source
alirezarezvani
License
MIT
Pinned revision
e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read alirezarezvani--grants at revision e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 4

references/reporter_post_patterns.md

RePORTER POST Patterns + Plan-Tier Detection

This reference answers exactly one decision: how does the grants skill query NIH RePORTER, and what plan-tier signals does it detect from Consensus responses?

The Critical Constraint

NIH RePORTER's API v2 is POST-only. web_fetch (which performs GET requests) will not work. You MUST use bash_tool + curl.

This is the #1 anti-pattern for the grants skill. If a future maintainer "simplifies" to web_fetch, RePORTER queries silently fail and the skill produces hollow institute-mapping sections.

RePORTER API Reference

  • Endpoint: https://api.reporter.nih.gov/v2/projects/search
  • Method: POST
  • Content-Type: application/json
  • No auth required for public-data queries
  • Rate limit: documented as 1 q/sec; the skill applies 1+ sec sequential pause per research-pack convention

Standard POST Templates

Narrow (AND) — 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": "deep learning electronic health records sepsis prediction"
      }
    },
    "limit": 50,
    "offset": 0,
    "include_fields": [
      "project_num",
      "project_title",
      "agency_ic_admin",
      "study_section",
      "fiscal_year",
      "principal_investigators",
      "abstract_text",
      "project_terms"
    ]
  }'

Broad (OR) — 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": "machine learning critical care sepsis early warning"
      }
    },
    "limit": 50
  }'

Dynamic Fiscal Year Window

NIH fiscal year runs Oct 1 → Sep 30. Current FY = year of next Sep 30.

Use scripts/fiscal_year_calculator.py:

python ../scripts/fiscal_year_calculator.py
# Output:
# Current calendar year: 2026
# Current fiscal year:   2026 (Oct 1 2025 - Sep 30 2026)
# Window (current + 3 prior): [2023, 2024, 2025, 2026]

Never hardcode years. A skill committed in 2025 with hardcoded [2022, 2023, 2024, 2025] produces stale results in 2027.

Institute Tally + Study Section Ranking

After both narrow + broad responses return, aggregate:

Institute tally

For each project: extract agency_ic_admin (the institute code like NCI, NHLBI, NIMH).

from collections import Counter
institute_counts = Counter()
for project in projects:
    institute_counts[project['agency_ic_admin']] += 1
top_institutes = institute_counts.most_common(3)

Surface in DOCX Section 3 as ranked table with project counts + brief institute mission.

Study section ranking

For each project: extract study_section.

study_section_counts = Counter()
for project in projects:
    section = project.get('study_section', '')
    if section:  # Some projects unassigned
        study_section_counts[section] += 1
top_sections = study_section_counts.most_common(2)

Surface in DOCX Section 6.

NOSI Discovery from RePORTER Results

NOSI (Notice of Special Interest) numbers appear as NOT-* in project abstracts, project terms, or related-FOA fields. Parse with regex:

import re
NOSI_RE = re.compile(r'NOT-[A-Z]{2,3}-\d{2}-\d{3}')
nosi_numbers = set()
for project in projects:
    abstract = project.get('abstract_text', '')
    nosi_numbers.update(NOSI_RE.findall(abstract))

For each NOSI number, fetch via web_fetch (NOSIs have predictable URLs):

https://grants.nih.gov/grants/guide/notice-files/{NOSI_NUMBER}.html

If fetch fails: log [NOSI {number} — fetch failed, not included]. Never fabricate NOSI details.

Plan-Tier Detection (Consensus)

Consensus has tiered plans with different per-query result caps. The skill detects from response text patterns:

| Pattern in response | Tier | Per-query cap | |---|---|---| | "Showing top 10" / "upgrade for more" | Free | 10 results | | Receives 20 results without "showing top" | Pro | 20 results | | Receives ≤3 results consistently | Unauthenticated / API quota issue | 3 results | | No response / 401 / 403 | Auth failure | n/a |

Surface at end of Phase 2A in DOCX audit log:

Plan tier detected: Free (Consensus returns ~10 results per query, capped). Total positioning landscape: 5 facets × 10 results = ~50 papers max. For deeper coverage, consider Consensus Pro (20/query).

This calibrates user expectations BEFORE they read the DOCX and wonder why coverage seems thin.

Sequential Execution Discipline

Per research-pack convention: 1 q/sec, never parallelize.

  • 5 Consensus searches (Phase 2A) sequential — pause 1+ sec between
  • 2 RePORTER POST searches (narrow + broad) sequential
  • N NOSI web_fetch calls sequential

Each call records timestamp via citation_tracker.py; second call within 1s is rejected.

Total Phase 2 wall-clock time: ~7-10 sec for searches + however long NOSI fetches take.

Error Handling

| Failure | Handling | |---|---| | Consensus 429 (rate limit) | Wait 3s, retry once, log to audit | | Consensus 0 results for a facet | Surface explicitly in DOCX positioning section; mark [no results — verify terminology] | | RePORTER 5xx | Retry once after 3s; if still failing, log and continue with what's available | | RePORTER <5 results on narrow | Document low count; rely on broad OR for coverage | | NOSI fetch fails | [NOSI {number} — fetch failed]; never fabricate | | 3 consecutive failures across tools | Halt; alert researcher with what's missing | | Auth failure (401/403) | Halt; tell user to check API key or MCP connection |

Citations (7 sources)

  1. NIH RePORTER API v2 documentation — https://api.reporter.nih.gov/documents/Data%20Element%20Descriptions.pdf. Authoritative spec for POST endpoint, field definitions, fiscal-year filter semantics. The skill's curl templates are direct applications.

  2. NIH Office of Extramural Research — NIH Guide for Grants and Contracts (https://grants.nih.gov/grants/guide). Source for NOSI / FOA URL structure. NOSI naming conventions (NOT-{IC}-{YY}-{NNN}) are documented here.

  3. praw library + Reddit API community guidance. Source for the "1 q/sec is the polite default" pattern that the skill applies to RePORTER even though RePORTER's documented limits are looser. Politeness with shared infrastructure.

  4. Mike Cohen, "Exponential Backoff and Jitter" — AWS Architecture Blog, 2015. Source for the "wait 3s + retry once" retry pattern. Research-workflow scale doesn't justify exponential backoff.

  5. curl documentation (https://curl.se/docs/manual.html). Source for POST body + Content-Type header syntax. The skill's curl templates follow curl --help.

  6. Maynez et al., "On Faithfulness and Factuality in Abstractive Summarization" — ACL 2020. Source for the source-discipline rule that justifies refusing to fabricate NOSI details when fetch fails. LLMs hallucinate plausible-looking NIH NOSI numbers; refuse.

  7. Susskind, D., "Show your work" — Communications of the ACM, 2024. Source for the audit-log section's role: transparent surfacing of what was queried, what was returned, what was cited. The audit-log table in DOCX Section 9 is this principle's implementation.