dongbeixiaohuo/writing-agent

✍️ Writing‑Agent v0.10.0 – An AI‑powered, stage‑driven writing pipeline

What it is – A locally‑run tool (CLI + optional Windows desktop app) that turns the act of writing a long‑form article into a repeatable, interrupt‑able workflow. It stitches together a series of stages (topic, stance, evidence ledger, outline, opening hook, draft, human‑izer, fact‑checker, reviewer, reader‑test, export) and stores the artefacts of each step on disk so you can pause, edit, or replay the process later.

Why it matters – Most AI‑writing utilities spit out a single draft and leave you with a black‑box output. Writing‑Agent makes the process transparent and auditable:

• Evidence‑backed – every factual claim is linked to an entry in 02_evidence_ledger.json and re‑checked before the final _clean.txt is emitted.
• Human‑in‑the‑loop – a Humanizer stage removes “AI‑flavour” while preserving only sourced material.
• Version‑safe – SHA‑256 hashes of the title and body are recorded in run_manifest.json; any change invalidates the previous run.
• Export options – plain‑text _clean.txt plus optional WeChat‑style HTML.

Key components (as laid out in the repo)

Directory / file	Role
writing-agent-app/	Source for the Windows preview desktop app (installer available in the releases).
.claude/skills/	Prompt‑templates and “skills” that drive each stage (e.g., research‑expert, opening‑tournament).
.claude/agents/	Agent definitions that wrap the chosen LLM (DeepSeek‑V3.2, 智谱 GLM, MiniMax).
.claude/workflows/	The ordered pipeline that the CLI executes.
scripts/	Helper scripts (node & python) for dependency checks, runtime contracts, and CI.
demo/工资的一半，是你受的气折算的/	A full end‑to‑end example showing every intermediate file the pipeline creates.
02_evidence_ledger.json	Structured ledger of facts (id, source, value) used for later fact‑checking.
run_manifest.json	Manifest that records hashes of the title, body, and the exact stage outputs for reproducibility.
style_registry.json & .claude/styles/	Library of verified writing styles; each style is marked verified or legacy_unverified.

Supported models – The tool does not require Claude’s own API; it works with any Anthropic‑compatible endpoint. The README recommends three models that have been tested end‑to‑end:

• DeepSeek‑V3.2 – cheapest entry point, good for a first run.
• 智谱 GLM – same performance, useful if you already have a GLM subscription.
• MiniMax – also tested, interchangeable.

Installation paths

1. Desktop preview (Windows) – download the Writing.Agent.App_0.1.0_x64-setup.exe from the Releases page and run it.
2. Full source install –

   git clone https://github.com/dongbeixiaohuo/writing-agent.git
   cd writing-agent
   npm ci          # installs Node deps (Node ≥ 18.17)
   # Python 3.11+ is also required for runtime scripts
   claude          # start Claude Code in the repo root
   

3. Plugin mode – add the repo to Claude’s marketplace and install the writing-agent plugin; the plugin will auto‑bootstrap the minimal folder structure (articles/, .claude/styles/, …) in any working directory.

Typical workflow (high‑level, as described in the README)

1. Stage 1 – Theme – write 01_theme.md to set the topic and constraints.
2. Stage 1b – Position – lock the article’s “teeth” in 01b_position.md.
3. Stage 2 – Evidence ledger – collect facts into 02_evidence_ledger.json.
4. Stage 3 – Outline – generate 03_outline.md.
5. Stage 5c – Opening hook – race several hooks and pick the best.
6. Stage 6 – Draft – produce draft_v1.md.
7. Stage 7 – Humanizer – strip AI‑specific phrasing → draft_v1_humanized.md.
8. Stage 10.5 – Fact‑checker – compare claims against the ledger; if any red‑flag appears the pipeline stops.
9. Stage 11 – Review / Pre‑publish – editorial review (pre_publish_review.md).
10. Stage 12 – Reader test – simulate platform‑specific engagement (wechat_reader_test.md).
11. Stage 13 – Export – final clean text _clean.txt and optional HTML for WeChat.
12. Stage 14 (optional) – Metrics – record exposure / share stats in publication_metrics.jsonl.

Why you might use it

• You write newsletters, opinion pieces, or industry reports and need audit‑ready drafts.
• You want a repeatable pipeline that can be paused and resumed without losing context.
• You care about fact‑checking and want the system to block publication if any claim lacks a source.
• You like to experiment with different LLM back‑ends without rewriting prompts.

License – MIT (see LICENSE).

---

Quick start for a newcomer

1. Install Node 18+, Python 3.11+, and Claude Code (via the native installer, Homebrew, or WinGet).
2. Clone the repo and run npm ci.
3. Start Claude (claude) in the repo root.
4. Prompt it:

   先别写新文章，先解释 demo/工资的一半，是你受的气折算的 里每个阶段文件各自起什么作用。
   

   The assistant will walk you through the demo’s files, confirming that the workflow is correctly loaded.

---

Bottom line – Writing‑Agent is a genuine, open‑source AI‑assisted writing system that treats article creation as a multi‑stage, verifiable pipeline rather than a one‑shot generation. It is fully configurable, works with several third‑party LLMs, and ships both as a CLI tool and a Windows desktop preview app.