Grant Proposal

Grant Proposal: From Research Ideas to Fundable Application

Draft a grant proposal based on: $ARGUMENTS

Overview

This skill turns validated research ideas into a structured, reviewer-ready grant proposal. It chains sub-skills into a grant-specific pipeline:

/research-lit → /novelty-check → [structure design] → [draft] → /research-review → [revise] → GRANT_PROPOSAL.md
  (survey)      (verify gap)     (aims + matrix)     (prose)    (panel review)     (fix)      (done!)

This is a parallel branch, not part of the linear Workflow 1→1.5→2→3 pipeline. After /idea-discovery produces validated ideas, the user can either:

• Go to /experiment-bridge → /auto-review-loop → /paper-writing (implement & publish)
• Go to /grant-proposal (write funding application first, then implement after funding)

                    ┌→ /experiment-bridge → /auto-review-loop → /paper-writing  (publish track)
/idea-discovery ────┤
                    └→ /grant-proposal → [get funded] → /experiment-bridge → ...  (funding track)

Grant proposals argue for future work (feasibility + potential), not completed work (results + claims). This skill handles the unique requirements of grant writing: narrative arc design, reviewer-facing structure, budget justification, timeline planning, and agency-specific formatting.

Constants

• GRANT_TYPE = KAKENHI — Default grant type. Supported: KAKENHI, NSF, NSFC, ERC, DFG, SNSF, ARC, NWO, GENERIC. Override via argument (e.g., /grant-proposal "topic — NSF").
• GRANT_SUBTYPE = auto — Sub-type within the grant agency. Examples: KAKENHI Start-up/Wakate/Kiban-B; NSFC Youth/Excellent-Youth/Distinguished/Overseas/Key; NSF CAREER/CRII/Standard. Auto-detected from argument or defaults to the most common sub-type.
• REVIEWER_MODEL = gpt-5.4 — Model used via a secondary Codex agent for proposal review. Must be an OpenAI model (e.g., gpt-5.4, o3, gpt-4o).
• OUTPUT_FORMAT = markdown — Output format. Supported: markdown, latex. LaTeX uses grant-specific templates when available.
• MAX_REVIEW_ROUNDS = 2 — Maximum external review-revise cycles before finalizing.
• OUTPUT_DIR = grant-proposal/ — Directory for generated proposal files.
• LANGUAGE = auto — Output language. Auto-detected from grant type: KAKENHI→Japanese, NSF→English, NSFC→Chinese, ERC→English, DFG→English (or German), SNSF→English, ARC→English, NWO→English. Override explicitly if needed.
• AUTO_PROCEED = false — At each checkpoint, always wait for explicit user confirmation before proceeding. Grant proposals require PI-specific judgment at every stage. Set true only if user explicitly requests fully autonomous mode.

💡 These are defaults. Override by telling the skill, e.g., /grant-proposal "topic — NSF CAREER, latex output" or /grant-proposal "topic — NSFC Youth, language: English".

Grant Type Specifications

KAKENHI (Japan — JSPS)

NSF (US)

NSFC (China — 国家自然科学基金)

ERC (EU — European Research Council)

DFG (Germany — Deutsche Forschungsgemeinschaft)

SNSF (Switzerland — Swiss National Science Foundation)

ARC (Australia — Australian Research Council)

NWO (Netherlands — Dutch Research Council)

GENERIC

For any grant not listed above. User provides section names, page limits, and review criteria via argument:

/grant-proposal "topic — GENERIC, sections: Background|Methods|Impact, language: English"

State Persistence (Compact Recovery)

Grant proposal drafting is a long task that may trigger context compaction. Persist state to grant-proposal/GRANT_STATE.json after each phase:

{
  "phase": 2,
  "grant_type": "KAKENHI",
  "grant_subtype": "Start-up",
  "language": "Japanese",
  "agent_id": "019cfcf4-...",
  "gap_statement": "...",
  "aims_count": 3,
  "status": "in_progress",
  "timestamp": "2026-03-18T15:00:00"
}

Write this file at the end of every phase. On invocation, check for this file:

• If absent or status: "completed" → fresh start
• If status: "in_progress" and within 24h → resume from saved phase (read GRANT_PROPOSAL.md and GRANT_REVIEW.md to restore context)
• If older than 24h → fresh start (stale state)

On completion, set "status": "completed".

Workflow

Phase 0: Input Parsing & Context Gathering

Parse $ARGUMENTS to extract:

1. Research direction/idea — may reference existing files or be a freeform description
2. Grant type — detect from keywords (e.g., "科研費"→KAKENHI, "NSF"→NSF, "国自然"→NSFC, "基金"→NSFC)
3. Grant sub-type — detect from keywords (e.g., "Start-up", "若手", "青年", "CAREER", "优青", "海外优青")
4. Overrides — output format, language, review rounds

Then gather context from the project directory:

1. Read IDEA_REPORT.md if it exists (from /idea-discovery)
2. Read refine-logs/FINAL_PROPOSAL.md if it exists (from /research-refine)
3. Read refine-logs/EXPERIMENT_PLAN.md if it exists (from /experiment-plan)
4. Read AUTO_REVIEW.md if it exists (from /auto-review-loop — prior review feedback is gold for grants)
5. Read NARRATIVE_REPORT.md or STORY.md if they exist
6. Read any existing literature notes or survey documents
7. Scan for the user's publication list (e.g., publications.md, cv.md, bio.md, CV.pdf)
8. Check for grant-proposal/GRANT_STATE.json (resume from prior interrupted run)

If insufficient context exists:

• No research idea at all → suggest running /idea-discovery first
• No literature survey → will invoke /research-lit inline in Phase 1
• No publication list → leave PI qualification section with [TODO: Add publications] placeholders
• Has AUTO_REVIEW.md → extract reviewer feedback and use it to strengthen the feasibility narrative

Phase 1: Literature & Landscape Positioning

Invoke /research-lit to ground the proposal in real literature, then search for competing funded projects:

/research-lit "$ARGUMENTS"

What this does:

• Reuse existing surveys if /research-lit was already run and notes exist
• Otherwise invoke /research-lit for multi-source literature search (arXiv, Scholar, Zotero, local PDFs)
• Search for funded projects in the same area via WebSearch:
  • KAKENHI → KAKEN database (https://kaken.nii.ac.jp/)
  • NSF → NSF Award Search (https://www.nsf.gov/awardsearch/)
  • NSFC → NSFC funded projects
  • Other agencies → general web search
• Identify competing groups and their recent publications
• Run /novelty-check on the proposed research direction to verify the gap is real:

  /novelty-check "[proposed gap statement]"
  

• Build the gap statement — the single most important sentence in the proposal:

  "Despite progress in [X], [specific gap] remains unaddressed because [reason].
  This proposal addresses this by [approach], which will [expected impact]."
  

🚦 Checkpoint: Present the landscape summary and gap statement to the user:

📚 Literature & landscape analysis complete:
- [key findings from literature]
- [competing funded projects found]
- Gap statement: "[the gap statement]"

Does this accurately capture the positioning? Should I adjust before designing the proposal structure?

⛔ STOP HERE and wait for user response. Do NOT auto-proceed unless AUTO_PROCEED=true was explicitly set by the user.

Options for the user:

• Reply "go" or "ok" → proceed to Phase 2 with current positioning
• Reply with adjustments (e.g., "focus more on X", "the gap should emphasize Y") → refine and re-present
• Reply "stop" → end the skill, save current progress to grant-proposal/DRAFT_NOTES.md

State: Write GRANT_STATE.json with phase: 1 and the gap statement.

Phase 2: Narrative Structure & Aims Design

Design the proposal's logical architecture before writing any prose.

2.1 Define Specific Aims (2-4)

Each aim must satisfy:

• Independently valuable — if one aim fails, others still produce publishable results
• Logically connected — Aim 1 enables Aim 2, Aim 2 informs Aim 3
• Concrete deliverables — each aim maps to specific outputs (papers, datasets, tools, benchmarks)
• Feasible within budget and timeline

2.2 Build Claims-Aims-Evidence Matrix

| Aim | Key Claim | Preliminary Evidence | Proposed Validation | Risk Level | Deliverable |
|-----|-----------|---------------------|--------------------|-----------:|-------------|
| Aim 1 | [claim] | [pilot data, prior work] | [experiments] | LOW | [paper, dataset] |
| Aim 2 | [claim] | [theoretical basis] | [experiments] | MEDIUM | [paper, tool] |

2.3 Design the Narrative Arc

Grant proposals follow a fundamentally different arc from papers:

Problem → Why Now → What We Propose → Why It Will Work → What We Will Deliver
         (not: Problem → Method → Results → Implications)

• Problem: What gap exists and why it matters (scientific + societal)
• Why Now: What recent developments make this the right time (new data, new methods, new need)
• What We Propose: The specific aims and approach
• Why It Will Work: Preliminary data, PI track record, team expertise, feasibility arguments
• What We Will Deliver: Concrete outputs, timeline, expected publications

2.4 Timeline & Milestones

Design year-by-year (or quarter-by-quarter) plan:

### Year 1
- Q1-Q2: [Aim 1 tasks]
- Q3-Q4: [Aim 1 completion + Aim 2 start]
- Expected outputs: [papers, datasets]

### Year 2
- Q1-Q2: [Aim 2 completion + Aim 3]
- Q3-Q4: [Aim 3 completion + synthesis]
- Expected outputs: [papers, tools, final report]

2.5 Structural Review

Invoke /research-review to get critical feedback on the proposal structure before drafting:

/research-review "[GRANT_TYPE] [GRANT_SUBTYPE] proposal structure:
Gap: [gap statement]
Aims: [aims list with claims-evidence matrix]
Timeline: [timeline]
— reviewer persona: [GRANT_TYPE] review panelist"

What this does:

• GPT-5.4 xhigh acts as a grant review panelist (not a paper reviewer)
• Evaluates aims independence, narrative arc, risk identification, timeline realism
• Identifies the single biggest reviewer concern
• Provides actionable fixes ranked by severity

Apply structural feedback before proceeding to drafting.

🚦 Checkpoint: Present the proposal structure to the user:

🏗️ Proposal structure designed:
- Gap: [gap statement]
- Aim 1: [title] — Risk: LOW
- Aim 2: [title] — Risk: MEDIUM
- Aim 3: [title] — Risk: LOW
- Timeline: [summary]
- Reviewer feedback: [key points from GPT-5.4]

Proceed to section drafting? Or adjust the structure?

⛔ STOP HERE. This is the most critical checkpoint — the proposal structure determines everything downstream.

Options for the user:

• Reply "go" or "ok" → proceed to Phase 3 (section drafting)
• Reply with structural changes (e.g., "merge Aim 2 and 3", "add an aim about X", "reduce to 2 aims") → redesign and re-present
• Reply "back" → return to Phase 1 to adjust the gap/positioning
• Reply "stop" → save current structure to grant-proposal/DRAFT_NOTES.md

State: Write GRANT_STATE.json with phase: 2, aims summary, and the reviewer agent id.

Phase 3: Section Drafting

Draft each section according to the grant type template. Write complete prose, not outlines or placeholders.

What this does:

• Writes all required sections in the agency-specific language and tone
• Pulls content from IDEA_REPORT.md, FINAL_PROPOSAL.md, and literature notes
• Uses /paper-illustration for figure generation (if user requests)
• Leaves [TODO] only for PI-specific information, [AMOUNT] for budget figures
• Outputs grant-proposal/GRANT_PROPOSAL.md

Drafting Order (optimized for narrative coherence)

1. Specific Aims / Research Objective — the "abstract" of the grant. Write first, refine last.
2. Background / Significance / State of the Art — establish the problem and gap.
3. Research Plan / Methods — per aim, with feasibility arguments.
4. Figures — generate key diagrams (see below).
5. Timeline & Milestones — year-by-year deliverables.
6. PI Qualification / Preparation Status — track record, team, infrastructure.
7. Budget Justification — narrative only (leave dollar/yen amounts as [AMOUNT] placeholders).
8. Broader Impacts / Societal Significance — if required by the grant type.

Figure Generation

Grant proposals benefit greatly from clear diagrams. Generate the following figures using SVG or matplotlib (save to grant-proposal/figures/):

1. 全体構成図 / Overview Diagram — Show the relationship between aims (Aim 1 → Aim 2 → Aim 3), shared resources (participants, stimuli, pipeline), and outputs. This is the single most important figure.
2. 実験パラダイム図 / Experimental Paradigm — Visual schematic of each paradigm (stimulus timing, conditions, EEG recording).
3. 年次計画 / Timeline Gantt Chart — Year-by-year (or H1/H2) milestones with deliverables.

For AI-generated publication-quality figures, invoke /paper-illustration:

/paper-illustration "Overview diagram showing [aims relationship + shared resources] for grant proposal"

For simpler diagrams (flowcharts, Gantt charts), generate clean SVG or matplotlib directly via code.

🚦 Figure Checkpoint: Before generating, ask which figures the user wants:

🎨 The following figures would strengthen this proposal:
1. 全体構成図 / Overview — aims relationship + shared resources
2. 実験パラダイム図 / Paradigm — stimulus timing + conditions
3. 年次計画 / Gantt — timeline with milestones

Which should I generate? (e.g., "1 and 3", "all", "skip")

⛔ Wait for user response. Generate only the requested figures.

Grant-Specific Drafting Guidelines

KAKENHI:

• Write in formal Japanese academic style (である調, not です/ます調)
• Use 「」for Japanese quotations, bold for emphasis
• Structure: 研究の学術的背景 → 研究期間内に何をどこまで明らかにするか → 本研究の学術的な特色・独創性
• Include explicit 年次計画 (yearly plan) with concrete milestones
• Emphasize 社会的意義 (societal significance)
• Reference related KAKEN-funded projects to show awareness of the field

NSF:

• Write in clear, direct English
• Use Aim-based structure with bold headings
• Preliminary data paragraphs for each Aim (with figure references)
• Broader Impacts must be concrete: specific outreach activities, broadening participation plans
• Include Results from Prior Support (if PI has prior NSF funding)

NSFC:

• Write in formal Chinese academic style
• 立项依据 must position work at 国际前沿 (international frontier)
• 创新性 section must list numbered innovation points (创新点)
• 研究基础 must cite PI's own publications (with IF and citations if possible)
• 可行性分析 must address: technical feasibility, team capability, time feasibility, equipment/conditions

ERC:

• Write a compelling "high-risk/high-gain" narrative
• Extended Synopsis must be self-contained and compelling
• Include Work Package table with deliverables and milestones
• Gantt chart (describe in text, or generate as figure)

For Each Section

1. Pull relevant content from IDEA_REPORT.md, FINAL_PROPOSAL.md, literature notes
2. Write complete prose — no [TODO] except for PI-specific information
3. Include figure/table placeholders where appropriate (e.g., [Figure 1: System architecture])
4. Cite references properly — use citation keys, will build bibliography later
5. Match the agency's tone and style — formal Japanese for KAKENHI, direct English for NSF, etc.

Phase 4: External Review

Invoke /research-review on the complete draft for grant-type-specific evaluation:

/research-review "Complete [GRANT_TYPE] [GRANT_SUBTYPE] proposal draft. Evaluate as a [GRANT_TYPE] review panelist using official criteria. [PASTE FULL PROPOSAL TEXT]"

What this does:

• GPT-5.4 xhigh acts as a grant review panelist
• Scores each section 1-5 using agency-specific criteria
• Identifies fatal flaws and recommends funding/revisions/rejection
• Provides ranked action items for improvement
• All feedback saved to grant-proposal/GRANT_REVIEW.md

⚠️ External review fallback: If reviewer agents are unavailable, skip external review. Note "External review skipped — no reviewer agent available. Consider running /auto-review-loop-llm separately." in GRANT_REVIEW.md. The proposal is still usable without external review.

If /research-review is invoked (preferred), it handles the external review internally. If you run the reviewer directly, use spawn_agent for Round 1 and send_input for follow-up rounds.

Round 1 (full draft review):

spawn_agent:
  reasoning_effort: xhigh
  message: |
    Review this complete [GRANT_TYPE] [GRANT_SUBTYPE] proposal draft.

    Act as a [GRANT_TYPE] review panelist. Evaluate using the official criteria:

    [INSERT GRANT-TYPE-SPECIFIC CRITERIA — see Grant Type Specifications above]

    For each section:
    1. Score 1-5 (5 = excellent)
    2. Strongest aspect
    3. Most critical weakness
    4. Specific fix suggestion (actionable, not vague)

    Overall assessment:
    - Would you recommend funding? (Yes / Yes with revisions / No)
    - Single most impactful change to improve funding chances?
    - Any fatal flaws?

    [PASTE FULL PROPOSAL TEXT]

Save the returned reviewer agent id in GRANT_STATE.json if you want to continue the same dialogue in Round 2.

Round 2+ (after revisions):

If MAX_REVIEW_ROUNDS > 1 and revisions were applied:

send_input:
  agent_id: [saved from Round 1]
  message: |
    [Round N review of revised [GRANT_TYPE] [GRANT_SUBTYPE] proposal]

    Since your last review, I have applied the following changes:
    1. [Change 1]: [what was done]
    2. [Change 2]: [what was done]
    3. [Change 3]: [what was done]

    Please re-evaluate. Same format: section scores, overall assessment, remaining weaknesses.
    Focus on whether the CRITICAL and MAJOR issues from Round 1 have been adequately addressed.

    [PASTE REVISED PROPOSAL TEXT]

Phase 5: Revision & Output

5.1 Apply Reviewer Feedback

Parse reviewer feedback into severity levels:

• CRITICAL — fatal flaws that would lead to rejection. Fix immediately.
• MAJOR — significant weaknesses. Fix before submission.
• MINOR — suggestions for improvement. Fix if time allows.

Implement CRITICAL and MAJOR fixes. If MAX_REVIEW_ROUNDS > 1, re-submit for another round via send_input.

5.2 Generate Output

Markdown output (default):

grant-proposal/
├── GRANT_PROPOSAL.md          # Complete proposal, all sections
├── GRANT_REVIEW.md            # Review history and reviewer feedback
├── GRANT_STATE.json           # State persistence file
├── figures/                   # Generated diagrams (if any)
└── references.bib             # Bibliography (if citations were used)

LaTeX output (when OUTPUT_FORMAT = latex):

grant-proposal/
├── main.tex                   # Master file
├── sections/
│   ├── aims.tex               # Specific Aims / Research Objective
│   ├── background.tex         # Background / Significance
│   ├── research_plan.tex      # Research Plan / Methods
│   ├── timeline.tex           # Timeline & Milestones
│   ├── pi_qualification.tex   # PI Qualification / Track Record
│   └── budget.tex             # Budget Justification (if applicable)
├── references.bib
└── figures/                   # Any generated diagrams

5.3 Final Checks

Before declaring done:

• All sections required by the grant type are present and complete
• Gap statement is clear and appears early in the proposal
• Each aim is independently valuable and logically connected
• Timeline includes concrete yearly milestones and deliverables
• PI qualification section has content (or clear [TODO] placeholders)
• Budget justification uses [AMOUNT] placeholders (no fabricated numbers)
• Language matches the grant type (Japanese for KAKENHI, Chinese for NSFC, etc.)
• No leftover [TODO] markers except for PI-specific information
• References are real (no hallucinated citations)
• Review feedback has been addressed (CRITICAL and MAJOR items)

🚦 Final Checkpoint: Present the completed proposal summary:

📝 Grant proposal draft complete:
- Type: [GRANT_TYPE] [GRANT_SUBTYPE]
- Language: [language]
- Aims: [N] aims covering [summary]
- Timeline: [N] years
- Review score: [summary from GPT-5.4]
- Output: grant-proposal/GRANT_PROPOSAL.md

Files saved to grant-proposal/. Please review and customize:
1. PI qualification section (add your publications and track record)
2. Budget amounts (replace [AMOUNT] placeholders)
3. Any [TODO] markers for personal information

What would you like to do next?
- "figures" → generate proposal diagrams
- "review again" → run another round of external review
- "latex" → convert to LaTeX format
- "done" → finalize

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.

• Do NOT fabricate budget amounts. Generate narrative budget justification only. Leave specific dollar/yen/yuan/euro amounts as [AMOUNT] placeholders for the user to fill in.

• Do NOT fabricate PI information. If no publication list is available, leave [TODO: Add publications] placeholders. Never invent papers, grants, or credentials.

• Do NOT hallucinate citations. Use references from literature survey. Mark uncertain citations with [VERIFY].

• Grant ≠ paper. A grant argues for future work (feasibility + potential). A paper argues for completed work (results + claims). Write accordingly — emphasize "what we will do" and "why it will work", not "what we found."

• Aims must be independently valuable. If Aim 2 fails, Aim 1 and Aim 3 should still produce publishable results.

• Preliminary data de-risks. Include any pilot results, existing datasets, or prior publications that demonstrate feasibility.

• Reviewer-facing structure. Bold key sentences. Use numbered lists for clarity. Make the reviewer's job easy.

• Cultural norms matter. KAKENHI expects 社会的意義; NSF expects Broader Impacts; NSFC expects 国际前沿 positioning. Missing these is a red flag for reviewers.

• Feishu notifications are optional. If ~/.codex/feishu.json exists, send checkpoint at each phase transition and pipeline_done at final output. If absent, skip silently.

Parameter Pass-Through

Parameters can be passed inline with — separator. They flow to sub-skills when invoked:

/grant-proposal "topic — KAKENHI Start-up, sources: zotero, arxiv download: true"

Composing with Other Skills

Sub-skills used by this skill

Funding Track (this skill's primary use case)

/idea-discovery "direction"              ← Workflow 1: find validated ideas
/research-refine "idea"                  ← sharpen the method
/grant-proposal "idea — KAKENHI"         ← this skill: write the grant proposal
                                          ← [submit & get funded]
/experiment-bridge                       ← implement experiments with funding
/auto-review-loop "results"              ← Workflow 2: iterate until submission-ready
/paper-writing                           ← Workflow 3: write the paper

Publish Track (skip this skill)

/idea-discovery → /experiment-bridge → /auto-review-loop → /paper-writing → submit