Fund-Application-Workflow

基金申报全流程协同写作 Skill — An AI coding assistant skill for planning, researching, structuring, drafting, reviewing, and revising Chinese grant applications and research proposals.

Overview

This is an AI coding assistant custom skill that guides researchers through the full lifecycle of preparing a formal grant application (基金申报书), project proposal (项目申请书), or research plan (研究计划书).

The skill is not a one-shot proposal generator. Instead, it acts as a top-level orchestrator that progressively transforms scattered ideas into a defensible, evidence-backed project draft.

Supports GitHub Copilot, Cursor, Claude Code, and Codex (OpenAI).

Features

• S1 — Project Card Builder: Structures raw ideas into a project_card with known/missing/inferred information
• S2 — Gap Clarifier: Identifies which gaps need user input vs. external research
• S3 — Outline Generator: Builds a section blueprint with draft-readiness assessment
• S4 — Goal & Content Expander: Converts vague goals into structured objectives, task modules, and integrated paragraphs; supports Mermaid module-relation diagrams
• S5 — Innovation Extractor: Derives defensible innovation claims with risk alerts and comparative evidence needs
• S6 — Review Simulator: Simulates multi-perspective peer review with actionable revision priorities
• S7 — Prose Drafter: Generates publication-ready Chinese prose for each section based on accumulated intermediates; auto-generates Mermaid flowcharts and Gantt charts where appropriate

Core Principles

1. Evidence First — Never fabricates facts, citations, data, or policy references; unsupported claims are marked [待补引用]
2. Prose First, Structure Second — Responds in Chinese prose by default; structured data (project cards, frameworks) are presented as Markdown tables/lists; pure JSON mode available on request
3. Tri-Route Research — Classifies every information gap as internal (user's existing sources), external (new literature needed), or either (verify internally first); not tied to any specific tool
4. Bilingual Research Strategy — Research prompts instruct the model to think and search in English for broader coverage, then respond in Chinese with key terms preserved in English

Installation

Automated (recommended)

Use the bundled install.py to auto-detect your environment and set up the skill:

python .github/skills/fund-application-workflow/install.py

The installer supports:

Platform	What it does
GitHub Copilot	Copies skill to .github/skills/ (default location)
Cursor	Creates .cursor/rules/fund-application-workflow.mdc
Claude Code	Appends skill reference to CLAUDE.md
Codex (OpenAI)	Appends skill reference to AGENTS.md

Options: --dry-run (preview only), --uninstall (remove).

Manual

Copy the .github/skills/fund-application-workflow/ directory into your workspace:

your-workspace/
└── .github/
    └── skills/
        └── fund-application-workflow/
            ├── SKILL.md
            ├── install.py
            ├── references/
            │   ├── prompt-templates.md
            │   ├── schemas.md
            │   └── subskills.md
            └── evals/
                ├── BENCHMARK.md
                ├── evals.json
                └── manual-review/
                    └── *.md

State Persistence

The skill saves project progress to .fund-workflow/state.json across sessions. Each stage incrementally updates the state file, so you never lose context between conversations.

Visualization

Supports Mermaid diagrams:

Diagram Type	Purpose	When
Flowchart	Technical roadmap & dependencies	S7 writing methodology sections
Graph	Module relationships	S4 expanding research content
Gantt	Timeline & milestones	S7 writing deliverables sections
Mindmap	Goal decomposition	S4 expanding objectives

Evaluation

The evals/ directory contains:

• evals.json — 7 structured evaluation scenarios with assertions and pass criteria
• BENCHMARK.md — Human-readable benchmark guide and evaluation checklist
• manual-review/ — Gold-standard manual review outputs for 6 of the 7 evals

Supported Task Types

Task Type	Route	Description
ideation	S1	Turn raw ideas into a project card
clarify_gaps	S2	Identify and triage information gaps
build_outline	S3	Generate section blueprint
expand_goals	S4	Expand goals and research content
extract_innovations	S5	Extract defensible innovation claims
review_draft	S6	Simulate peer review
draft_section	S7	Draft publication-ready prose by section
revise_after_review	S6 + backflow	Post-review revision planning
full_workflow	Auto	Automatically advance to the next needed stage

License

MIT