GSD-AW — Get Shit Done-Academic Writing

Spec-driven academic paper writing framework. Chain AI agents through a proven pipeline: question → research → methodology → plan → write → cite → figure → abstract → compile.

---

Languages: English · 简体中文 · 日本語 · Français

---

Install (one command)

npm install -g @2p1c/harness-writing

Restart your Claude Code session after install. All skills auto-discover.

---

Prerequisites

GSD-AW requires LaTeX and optionally markitdown. Install before first use.

LaTeX (required)

macOS:

brew install --cask mactex

Linux (Debian/Ubuntu):

sudo apt install texlive-latex-base latexmk

Windows (WSL2 recommended):

# In WSL2
sudo apt install texlive-latex-base latexmk

Or install TeX Live natively on Windows.

markitdown — PDF extraction (optional)

markitdown converts PDFs to clean Markdown for literature review.

macOS:

conda install -c conda-forge markitdown
# or
brew install --cask mambaforge && conda install -c conda-forge markitdown

Linux:

conda install -c conda-forge markitdown
# or
wget "https://github.com/daltonmatos/markitdown/releases/latest/download/markitdown-x86_64-linux.tar.gz" \
  && tar -xzf markitdown-x86_64-linux.tar.gz && sudo mv markitdown /usr/local/bin/

Windows:

# With conda/mamba
conda install -c conda-forge markitdown

# Or download from GitHub releases and add to PATH

---

Start

/aw-init

回答 5 类问题（研究问题、研究思路、方法论、约束条件、参考资料）→ 生成研究简报。

---

Full Pipeline

/aw-init              → Research Brief
       ↓
/aw-execute          → Write all sections (wave-parallel)
       ↓
/aw-cite             → Verify citations
       ↓
/aw-table            → Build tables (provide CSV)
/aw-figure           → Generate figures (PlantUML + matplotlib)
       ↓
/aw-abstract         → Write 250-word abstract
       ↓
/aw-finalize         → make paper — verify & ready

---

Phase Notes

/aw-init

• Deep questioning across 5 categories before writing starts
• Auto-detects prior Research Brief — ask to reuse or start fresh
• --quick flag: skip Discuss checkpoints, go straight to research

/aw-execute

• Reads ROADMAP → groups tasks into dependency-ordered waves
• Wave 1: tasks with no dependencies → parallel
• Wave 2+: tasks depending on prior waves → sequential per wave
• After each wave: quality gate checkpoint (continue / modify / pause)
• Manual compile: run make paper yourself when ready

/aw-cite

• Scans all \cite{} keys against references.bib
• Auto-fixes missing entries from literature.md
• Run after any section edit that adds citations

/aw-table

• Asks for CSV data per table (dataset, baseline, ablation, results)
• If no CSV yet → leaves \placeholder{tab:name} in section
• Auto-inserts \input{tables/{name}} at correct location

/aw-figure

• Pipeline diagrams → PlantUML .tex
• Result plots → Python matplotlib .pdf
• If data unavailable → \placeholder{fig:name}
• Auto-inserts \input{figures/fig-name} at correct location

/aw-abstract

• Reads all section drafts
• Synthesizes 250-word IMRAD abstract (Background→Objective→Method→Results→Conclusion)
• No citations in abstract
• Run after all sections complete

/aw-finalize

• make paper — full compile
• Checks: undefined refs, unresolved citations, word count, abstract present
• Updates STATE.md to "ready for submission"

---

Writing Order

Methodology → Results → Introduction → Discussion → Conclusion → Abstract

---

All Commands

Command	Phase	Description
/aw-init	Init	Start new paper — questioner → brief
/aw-execute	Phase 2	Execute wave plan — write all sections
/aw-cite	Phase 3	Verify citations resolve
/aw-table	Phase 3	Build tables from CSV
/aw-figure	Phase 3	Generate figures
/aw-abstract	Phase 3	Write abstract
/aw-finalize	Phase 3	Compile & verify
/aw-review	Any	Section quality review
/aw-wave-planner	Manual	Re-plan waves from ROADMAP
/aw-pause	Any	Save writing session (before break)
/aw-resume	Any	Resume from checkpoint

---

Workflow Quick Ref

/aw-init
    │
    ├── aw-questioner → research-brief.json
    ├── aw-discuss-1 → confirm brief
    ├── [research + methodology] → literature.md + methodology.md (parallel)
    ├── aw-discuss-2 → consistency check
    ├── aw-planner → ROADMAP.md + STATE.md
    └── aw-discuss-3 → approve plan

/aw-execute
    │
    ├── aw-wave-planner → wave-plan.md
    ├── Wave 1 (parallel) → aw-write-*
    ├── aw-review → quality gate
    ├── Wave 2 (parallel) → aw-write-*
    ├── ...
    └── phase merge → sections/{chapter}.tex

/aw-cite → /aw-table → /aw-figure → /aw-abstract → /aw-finalize

/aw-pause  →  Save checkpoint before break
/aw-resume →  Resume from checkpoint

---

Project Structure

manuscripts/
└── {paper-slug}/
    ├── project.yaml
    ├── references.bib
    ├── main.tex
    └── sections/
        ├── intro/
        │   ├── 1-1-background.tex
        │   └── ...
        ├── related-work/
        ├── methodology/
        ├── experiment/
        ├── results/
        ├── discussion/
        ├── conclusion/
        └── tables/

Paragraph files (sections/{chapter}/{task-id}.tex) are individual units — one per paragraph, merged by wave executor into chapter .tex files.

---

Troubleshooting

Issue	Fix
Undefined reference in log	Run /aw-cite to find missing keys
Citation [?]	make clean && make paper
elsarticle.cls not found	tlmgr install elsarticle
/aw-init skill not found	Restart Claude Code after npm install
markitdown not found	See install instructions above for your OS

---

Skills Count

24 skills across 3 phases + session management.

Phase	Skills
Phase 1 — Orchestration	aw-questioner, aw-discuss-1/2/3, aw-research, aw-methodology, aw-planner, aw-orchestrator
Phase 2 — Execution	aw-wave-planner, aw-execute, aw-review, aw-write-intro, aw-write-related, aw-write-methodology, aw-write-experiment, aw-write-results, aw-write-discussion, aw-write-conclusion
Phase 3 — Polish	aw-cite, aw-table, aw-figure, aw-abstract, aw-finalize
Session	aw-pause, aw-resume

---

Acknowledgments

This framework was built using the GSD (Get Shit Done) agent orchestration methodology. Learn more at https://github.com/gsd-build/get-shit-done.

License

MIT