Grant Proposal Skill

Core Philosophy

Three principles govern every interaction:

1. Grant applications are arguments, not requests. Every section must advance a persuasive case. The narrative arc is: problem is important, you are the right person, your approach will work, the investment is justified.
2. Write like a domain expert, not a template filler. Generic language kills proposals. Every sentence must reflect deep knowledge of the specific field.
3. Grant is not Paper. A paper reports results; a grant sells a future. Different narrative arc, different evidence standards, different rhetoric.

Additional operating principles:

• Reviewer perspective, not applicant perspective. Always ask: "What would a tired reviewer scanning 80 proposals think when reading this sentence?"
• Every claim needs evidence; every expense needs task traceability.
• Two-phase drafting model: internal planning (with numbered scaffolding) is always purged before producing final output. The user never sees S1/S2/S3/S4 markers or internal notes in deliverables.

---

Routing Logic

On first interaction, determine the track:

IF user mentions NSFC / 国自然 / 青年基金 / 面上 / 地区 / 重点 / Chinese agency
   → CN MODE
ELIF user mentions NSF / NIH / DOE / DARPA / NASA / R01 / R21 / CAREER / US agency
   → US MODE
ELSE
   → ASK: "Are you targeting a US agency (NSF, NIH, DOE, DARPA, NASA) or a
     Chinese agency (NSFC programs)? This determines the template, structure,
     and review criteria I will use."

Language strategy:

• CN mode: draft proposal content in Chinese (中文), but interact in whatever language the user uses.
• US mode: draft proposal content in English, interact in whatever language the user uses.
• Internal skill instructions are always in English.

---

State Persistence

All session state is saved to GRANT_STATE.json in the working directory.

GRANT_STATE.json Schema

{
  "meta": {
    "track": "US" | "CN",
    "agency": "NSF" | "NIH" | "DOE" | "DARPA" | "NASA" | "NSFC",
    "program": "string (e.g., CAREER, R01, 青年科学基金)",
    "created": "ISO-8601",
    "last_modified": "ISO-8601",
    "current_phase": "0"|"1"|"2"|"3"|"4"|"5"|"complete",
    "current_step": "string"
  },
  "profile": {
    "applicant_name": "",
    "institution": "",
    "career_stage": "early | mid | senior",
    "field": "",
    "subfield": "",
    "roi_score": 0-15,
    "recommended_programs": []
  },
  "structure": {
    "title": "",
    "claims_aims_evidence_matrix": [],
    "outline": {},
    "figure_plan": []
  },
  "drafts": {
    "section_name": {
      "version": 1,
      "status": "planning | drafting | polished | reviewed",
      "file_path": "",
      "backup_path": ""
    }
  },
  "review": {
    "tier1_results": {},
    "tier2_results": {},
    "severity_report": []
  },
  "simulated_review": {
    "scores": {},
    "weaknesses": [],
    "revision_suggestions": []
  }
}

Rules:

• Read GRANT_STATE.json at the start of every conversation turn to resume context.
• Write GRANT_STATE.json after completing any phase or significant sub-step.
• If the file does not exist, create it during Phase 0.

---

Safety Rules

1. Auto-backup before writes. Before overwriting any file, copy the existing version to backups/<section_name>_v<N>.<timestamp>.txt. Use Bash cp for this. If backups/ does not exist, create it with mkdir -p backups before the first backup.
2. Never modify the user's original files without confirmation. If the user provides source files, work on copies. Always ask before writing back.
3. Warn on destructive operations. If a phase would discard previous work (e.g., re-running Phase 1 after Phase 2 drafting), warn the user and require explicit confirmation.
4. Sensitive data. Never include PI personal information (SSN, bank details) in any generated file. If encountered, warn and redact.

---

Reference Files

The skill uses supporting files in sibling directories:

• references/us/ — US agency guidelines: nsf_guide.md, nih_guide.md, doe_guide.md, darpa_guide.md, nasa_guide.md
• references/cn/ — CN agency guidelines: nsfc_guide.md
• references/common/ — shared resources: reviewer_personas.md, common_mistakes.md, resubmission.md
• references/rubrics/ — scoring rubrics: nsf_rubric.json, nih_rubric.json, nsfc_rubric.json
• templates/us/ — US templates: nih_specific_aims.md, nsf_project_summary.md, budget_justification.md
• templates/cn/ — CN templates: nsfc_justification.md, nsfc_research_content.md, nsfc_research_foundation.md, nsfc_abstract_5sentence.md
• config.yaml — skill configuration: supported agencies/programs, golden ratio benchmarks, AI-flavor patterns, severity levels. Read at Phase 0 initialization.
• scripts/ — deterministic check scripts:
  • validate_length.py — section length vs golden ratio/page limits
  • validate_citations.py — citation consistency and completeness
  • compliance_check.py — format compliance and AI-flavor detection

When a phase requires a reference or template, load it with Read from these directories. If a needed file is missing, inform the user and proceed with built-in knowledge, noting the gap.

Lazy Loading: Do NOT read all reference files at once. Load only the files needed for the current phase and agency track. For example:

• Phase 1 (CN track): read references/cn/nsfc_guide.md only, not all US guides
• Phase 4 (NIH): read references/rubrics/nih_rubric.json + references/common/reviewer_personas.md, not NSF/NSFC rubrics
• Templates: read the specific template being used, not all templates This keeps context focused and reduces token usage by ~60%.

---

Phase 0: Project Profiling & Grant Matching

Entry Criteria

• User has initiated a conversation about a grant proposal.

Workflow

Step 0.1 — Collect Applicant Profile

Gather (ask if not provided):

• Name, institution, department
• Career stage: early-career (< 5 yrs post-PhD), mid-career, senior
• Research field and subfield
• Track record summary: key publications, prior funding, preliminary data
• For CN: age (relevant for Youth Fund 青年科学基金 age cap of 35/40)
• For US: citizenship/residency status (relevant for some programs)

Step 0.2 — Collect Project Concept

Gather:

• One-paragraph project description
• Key innovation / what is new
• Why now? (timeliness)
• Preliminary data available? (yes/no/partial)
• Target budget range
• Target submission deadline

Step 0.3 — ROI Scoring (0-15)

Score the project's fundability across five dimensions (0-3 each):

Dimension	0	1	2	3
Significance	Incremental	Moderate gap	Clear gap	Urgent national priority
Innovation	Standard method	Novel combination	New approach	Paradigm shift potential
Investigator fit	Tangential	Related	Strong match	World expert
Preliminary data	None	Conceptual	Partial	Convincing dataset
Timeliness	No urgency	Modest momentum	Active field	Hot topic + policy alignment

Report the total score and interpretation:

• 0-5: High risk. Recommend strengthening concept before applying.
• 6-9: Competitive with strong writing. Proceed with caveats noted.
• 10-12: Strong candidate. Proceed confidently.
• 13-15: Exceptional. Consider flagship programs.

Step 0.4 — Agency & Program Recommendation

Based on track, field, career stage, and ROI score, recommend 1-3 programs:

US Track Programs:

Agency	Program	Best For
NSF	CAREER	Early-career faculty, broad impact
NSF	Standard/Collaborative	Established investigators
NIH	R01	Biomedical, 4-5 year projects
NIH	R21	Exploratory/high-risk biomedical
DOE	Early Career	Energy/physics early-career
DARPA	Young Faculty Award	Defense-relevant, high-risk
NASA	FINESST	Graduate student fellowships

CN Track Programs (NSFC):

Program	Chinese Name	Best For
Youth Fund	青年科学基金	Under 35 (male) / 40 (female), first NSFC
General Program	面上项目	Established researchers, broad
Regional Fund	地区科学基金	Researchers at western/regional institutions
Key Program	重点项目	Senior PIs, larger scope

Present recommendation with reasoning. Get user confirmation before proceeding.

Step 0.5 — Initialize State

Create GRANT_STATE.json with profile, track, agency, program. Set current_phase: "1".

Exit Criteria

• GRANT_STATE.json exists with completed profile section.
• User has confirmed agency/program selection.

---

Phase 1: Structure Planning

Entry Criteria

• Phase 0 complete. GRANT_STATE.json has profile and agency/program.

Reference Loading

• Read references/us/nsf_guide.md or references/us/nih_guide.md (US track) or references/cn/nsfc_guide.md (CN track) depending on the selected agency.
• Read references/common/common_mistakes.md for pitfalls to avoid during planning.

Workflow

Step 1.1 — Title Crafting

Generate 3-5 candidate titles following agency conventions:

• US: Typically "Action-Oriented Noun Phrase: Specific Technical Approach"
  • NSF CAREER example: "CAREER: Enabling Scalable X Through Novel Y"
• CN: Typically "基于[方法]的[对象][目标]研究"
  • NSFC example: "基于深度学习的城市地表温度时空精细化反演研究"

User selects or modifies. Save to state.

Step 1.2 — Claims-Aims-Evidence Matrix

Build a matrix connecting the argument structure:

| Claim (Why it matters) | Aim/Objective | Key Evidence | Gap Addressed |
|------------------------|---------------|--------------|---------------|
| Claim 1: ...           | Aim 1: ...    | Prelim data, lit | Gap 1: ... |
| Claim 2: ...           | Aim 2: ...    | Method validation | Gap 2: ... |
| Claim 3: ...           | Aim 3: ...    | Pilot study  | Gap 3: ...    |

Rules:

• Every claim must have at least one piece of evidence.
• Every aim must address at least one gap.
• 2-4 aims is typical. More than 4 signals scope creep.
• Aims should be independent enough that failure of one does not block others.

Save matrix to state.

Step 1.3 — Outline Generation

US Track — Generate skeleton for:

For NIH R01/R21:

1. Specific Aims (1 page)
   • Opening paragraph: significance + gap
   • Long-term goal + objective of this application
   • Central hypothesis + rationale
   • Aim 1 with hypothesis and approach summary
   • Aim 2 with hypothesis and approach summary
   • Aim 3 (if applicable)
   • Payoff paragraph
2. Research Strategy
   • Significance (establish importance, identify gap, state contribution)
   • Innovation (conceptual, technical, methodological novelty)
   • Approach (per aim: rationale, methods, expected outcomes, pitfalls, alternatives, timeline)
3. Project Summary / Abstract

For NSF:

1. Project Summary (1 page: overview, intellectual merit, broader impacts)
2. Project Description (15 pages max)
   • Introduction + background
   • Proposed research (per aim)
   • Broader impacts
   • Results from prior support
   • Timeline / milestones
3. References Cited

CN Track — Generate skeleton for NSFC:

Page Budget (Golden Ratio): Cite these benchmarks explicitly when planning:

• 立项依据 ≈ 30% of total pages (including references; actual text ~4-6 pages)
• 研究内容+创新+年度计划 ≈ 50% (figure-heavy, 10-20 figures)
• 研究基础+工作条件 ≈ 20%
• Total target: 12,000-15,000 characters, 12-15 pages, under 28 pages hard limit

1. Title and basic info (项目名称、基本信息)
2. Project rationale (立项依据) — use the four-paragraph closure model:
   • Para 1: Field significance + macro context (大背景)
   • Para 2: Current state of research + what has been achieved (研究现状)
   • Para 3: Remaining problems + specific gaps (存在问题)
   • Para 4: This project's entry point + why it will work (本项目切入点) The four paragraphs must form a logical closure: significance → progress → gaps → your solution. The reader should feel "of course this is the next step" by paragraph 4.
3. Research content (研究内容) — internal planning uses S1-S4 structure. S1-S4 are planning DIMENSIONS, not timeline phases:
   • S1: Problem decomposition (问题分解) — break the core question into 3-4 researchable modules, each mapping to a research content section
   • S2: Feasibility pre-check (可行性预评估) — for each module, assess key technique maturity (high/medium/low), risk points, backup plans
   • S3: Dependency mapping (依赖关系) — which module outputs feed into which module inputs? What can run in parallel? Define milestones.
   • S4: Innovation audit (创新点验证) — for each claimed innovation, self-check: has anyone done similar work? Is it method-level or conceptual-level? Can it be stated in one clear sentence? IMPORTANT: S1-S4 markers are for internal planning ONLY. They are purged before producing any user-facing output. The final text flows as continuous prose organized by sub-topic headings. Do NOT present S1-S4 as Year 1/2/3/4.
4. Key scientific questions (拟解决的关键科学问题, 2-3 items)
5. Research plan and timeline (研究方案及可行性分析)
6. Innovation points (特色与创新之处, 2-3 bullet points)
7. Expected outcomes (预期研究成果)
8. Research foundation (研究基础与工作条件)
9. Budget justification (经费预算说明)

Step 1.4 — Figure Planning

Every proposal needs figures. Plan at minimum:

• 1 conceptual/overview figure (research framework or hypothesis model)
• 1 preliminary data figure (or technical approach diagram if no prelim data)

For each planned figure, note:

• Purpose (what argument does it support?)
• Placement (which section?)
• Data source (existing or to be created?)

Save figure plan to state.

Step 1.5 — Save & Checkpoint

Write full outline and matrix to GRANT_STATE.json. Set current_phase: "2". Summarize the plan to the user and ask for approval before moving to drafting.

Exit Criteria

• Outline approved by user.
• Claims-Aims-Evidence matrix complete.
• Figure plan documented.
• GRANT_STATE.json updated with structure section.

---

Phase 2: Section-by-Section Drafting

Entry Criteria

• Phase 1 complete. Outline approved. State file has structure.

Reference Loading

• Read the appropriate templates from templates/us/ (US track) or templates/cn/ (CN track) for the sections being drafted.
• Read references/common/common_mistakes.md for common drafting pitfalls.

General Drafting Protocol

For EVERY section, follow the two-phase model:

Planning Phase (internal, not shown to user as final output):

1. Identify the section's argumentative role in the overall proposal.
2. List the key points that must appear, with evidence for each.
3. Note the review criteria this section addresses.
4. Set target length based on agency page limits and golden-ratio benchmarks.
5. For CN: use S1-S4 internal numbering to organize thoughts.

Narrative Phase (user-facing output):

1. Write flowing, expert-level prose. No bullet lists in narrative sections unless the agency template calls for them.
2. Purge all internal planning markers (S1, S2, etc.).
3. Ensure every paragraph has a topic sentence and advances the argument.
4. Include figure references where planned.
5. Match the voice and tone conventions of the target agency.

Section-Specific Guidance

US Track: Specific Aims / Project Summary

The Specific Aims page is the most important page in any NIH proposal. Structure:

• Opening hook: one sentence establishing the big problem.
• Narrow to the specific gap (2-3 sentences with citations).
• "The long-term goal of [PI] is... The objective of this application is..."
• "Our central hypothesis is... This hypothesis is based on..."
• Aim 1: [action verb] [what] [method] [expected outcome]
• Aim 2: same pattern
• Aim 3: same pattern (optional)
• Payoff paragraph: what changes if this succeeds?

For NSF Project Summary: three separate sections clearly labeled Overview, Intellectual Merit, Broader Impacts. Each ~200 words. No jargon in Broader Impacts — a program officer outside your subfield will read it.

CN Track: Project Rationale (立项依据)

Follow the four-paragraph closure model from Step 1.3. Additional rules:

• Citation density: aim for 30-50 references. Under 20 signals shallow review.
• Include both international and domestic (Chinese) references.
• Do not merely list references — synthesize and critique.
• End with a clear statement: "因此，本项目拟..." connecting rationale to your proposed work.

CN Track: Research Content (研究内容)

Internal planning (S1-S4) guides the structure, but output is organized by research sub-topics. Each sub-topic section includes:

• What will be studied (研究对象)
• How it connects to the scientific question
• Methods to be used
• Expected results for this sub-topic

Agency-Specific Templates

Load the appropriate template from templates/ for the target agency/program. If a template exists, use it as the structural scaffold. Key templates:

• templates/us/nih_specific_aims.md — NIH Specific Aims page template
• templates/us/nsf_project_summary.md — NSF Project Summary template
• templates/us/budget_justification.md — US budget justification template
• templates/cn/nsfc_justification.md — NSFC project rationale (立项依据) template
• templates/cn/nsfc_research_content.md — NSFC research content (研究内容) template
• templates/cn/nsfc_abstract_5sentence.md — NSFC five-sentence abstract (五句模型) template

Review Criteria Alignment

While drafting each section, keep the relevant review criteria visible:

NIH (Scored Review Criteria):

• Significance, Investigator(s), Innovation, Approach, Environment

NSF (Merit Review Criteria):

• Intellectual Merit, Broader Impacts

NSFC (评审要点):

• 科学意义 (Scientific significance)
• 创新性 (Innovation)
• 研究方案可行性 (Feasibility of research plan)
• 研究基础 (Research foundation)

After drafting each section, do a self-check: "Does this section explicitly address the review criteria it should? If a reviewer is scoring criterion X, what in this section earns a high score?"

Figures

At least 1-2 figures are mandatory. When drafting reaches a section where a figure was planned:

1. Describe the figure in detail (what it shows, layout, labels).
2. If the user can provide the figure, request it.
3. If generating a conceptual diagram, describe it precisely so the user can create or commission it.
4. Insert a placeholder: [FIGURE X: description] in the draft.

Auto-Backup & Checkpoints

• Before writing any section draft to a file, back up the previous version: backups/<section_name>_v<N>.<timestamp>.txt
• After completing each section, update GRANT_STATE.json:
  • Set section status to "drafting" or "polished"
  • Increment version number
  • Record backup path
• After completing ALL sections for a major component (e.g., all of Research Strategy), pause and checkpoint: summarize what was written, ask user to review before proceeding.

Exit Criteria

• All sections drafted according to the outline.
• At least 1-2 figure placeholders inserted.
• Each section backed up and tracked in state.
• current_phase set to "3" in state.

---

Phase 3: Quality Review

Entry Criteria

• Phase 2 complete. All sections drafted.

Reference Loading

• Read references/common/common_mistakes.md for known quality issues to check.
• Read the agency guide (references/us/nsf_guide.md, references/us/nih_guide.md, or references/cn/nsfc_guide.md) to verify compliance requirements.

Tier 1: Deterministic Checks

Run scripts from the scripts/ directory for automated checks. If a script is not available, perform the check manually.

Length vs. Golden Ratio

• Check each section's length against agency page/word limits.
• Compare to golden-ratio benchmarks (e.g., for NIH R01 Research Strategy 12 pages: Significance ~2.5pp, Innovation ~1.5pp, Approach ~8pp).
• Flag sections that deviate more than 5% from benchmark ratios (matches scripts/validate_length.py threshold).

Citation Consistency

• Every in-text citation has a matching entry in the reference list.
• No orphaned references (listed but never cited).
• Citation format matches agency requirements (e.g., NIH uses numbered, NSF uses author-year typically).
• For CN: check that both Chinese and international references are included.

Format Compliance

• Font size, margins, page limits per agency specs.
• Required sections present (e.g., NSF requires Data Management Plan, Postdoctoral Mentoring Plan if applicable).
• Budget numbers consistent between narrative and budget forms.
• For CN: character count limits for abstract (400 characters), keywords (3-5).

Run checks using scripts if available:

python3 scripts/validate_length.py <proposal_dir> --mode cn|us --json
python3 scripts/validate_citations.py <file_or_dir> --mode cn|us --json
python3 scripts/compliance_check.py <file> --agency nsf|nih|nsfc --json

If scripts are not available or fail, perform these checks manually by reading the draft files and applying the rules from the agency guide. Document findings in the same P0/P1/P2 format regardless of check method.

Tier 2: AI Semantic Checks

Logic Coherence

• Read the full proposal start-to-finish.
• Check: Does the rationale logically lead to the proposed work?
• Check: Are aims independent but synergistic?
• Check: Do methods match objectives?
• Check: Does the timeline align with scope?
• Check: Does the budget align with the proposed activities?

AI-Flavor Detection (16-Item Checklist)

Scan the draft for these common AI-writing markers. Flag any found:

Read the full 24-item checklist from references/common/ai_flavor_checklist.md (items 1-16 for English, 17-24 for Chinese). For each flagged item, provide the specific location and a concrete revision.

Cross-Section Terminology Consistency

• Key terms, abbreviations, and acronyms are used consistently throughout.
• The same concept is not called different names in different sections.
• Abbreviations are defined at first use.

Severity Report

Classify every finding by severity:

• P0 (Critical): Will likely cause rejection. Must fix before submission. Examples: missing required section, exceeding page limit, contradictory aims.
• P1 (Major): Significantly weakens the proposal. Should fix. Examples: weak rationale, unclear methods, AI-flavor detected.
• P2 (Minor): Polish items. Fix if time permits. Examples: awkward phrasing, minor formatting, citation style inconsistency.

Present as a structured table:

| # | Severity | Section | Issue | Recommendation |
|---|----------|---------|-------|----------------|
| 1 | P0 | Specific Aims | Aim 3 overlaps with Aim 1 scope | Merge or differentiate |
| 2 | P1 | Significance | No quantitative impact data | Add statistics from ... |
| 3 | P2 | Approach | "Delve" used 4 times | Replace with varied verbs |

Save full report to GRANT_STATE.json review section.

Exit Criteria

• All Tier 1 checks run and results documented.
• All Tier 2 checks run and results documented.
• Severity report generated with P0/P1/P2 classifications.
• current_phase set to "4" in state.
• User has reviewed the report and decided which items to address.

---

Phase 4: Simulated Review

Entry Criteria

• Phase 3 complete. Quality issues addressed (at minimum all P0 items).

Reference Loading

• Read the appropriate rubric: references/rubrics/nsf_rubric.json, references/rubrics/nih_rubric.json, or references/rubrics/nsfc_rubric.json.
• Read references/common/reviewer_personas.md for detailed persona definitions and scoring guidance.
• If resubmission, also read references/common/resubmission.md.
• Always include AI-flavor detection as part of the simulated review — use the 24-item checklist from Phase 3 (items 1-16 for English, 17-24 for Chinese). This is a distinct value-add that reviewers increasingly notice.

US Track: Three-Pass Reviewer Simulation