Quickstart
From pip install writing-agent to a finished, exported article in one command.
Get started →
WRITING AGENT
An autonomous writing studio in your terminal. Give it an abstract; it plans, researches, writes, critiques itself, revises, and hands you a finished manuscript.
6export formats
3-stepwrite · critique · revise
0api key in fake mode
MITlicensed
$ writing-agent write —abstract “Why most RAG evaluation metrics mislead teams”
from-idea-to-better-rag-evals 01:32 elapsed · 23,976 tokens · $0.07
███████████░░░░░░░░░░░░░ 3/6
Section 4: Building an eval set your team will actually trust · ⠹ critiquing..
────────────────────────────────────────────────
✓ committed section 1 (+ summary)
✓ committed section 2 (+ summary)
✓ committed section 3 (+ summary)
· verdict=approve confidence=0.88 insight=4/5
The problem
Section titled “The problem”LLMs will happily write you 5,000 words on anything. Most of it is slop - generic structure, hedged claims, “delve” and “moreover”, fabricated citations, no thesis worth disagreeing with. Getting publishable long-form writing out of a model takes a dozen rounds of prompting, checking, and fixing - the parts that were supposed to be automated.
The solution
Section titled “The solution”WRITING AGENT is a pipeline, not a prompt. Every chapter and section runs through a write → critique → revise loop with a separate critic model that blocks AI tells, unsupported claims, thesis drift, and length misses - and it only escalates to you when it’s genuinely stuck (or never, in autonomous mode).
| What it does | How |
|---|---|
| Argues something | A contestable thesis with a steelmanned counterargument, enforced by the critic in every section |
| Sounds like a person | A deterministic AI-tell scanner + a surgical humanizer that rewrites only flagged sentences |
| Gets facts from somewhere | Optional web research per unit - shallow snippets or deep multi-source page fetches, with numbered citations renumbered to match the final references |
| Proves it | Every article ships an evidence report - the thesis it argues + every source ranked by influence (0–100), built from the finished piece with no model call |
| Remembers what it wrote | A durable on-disk canon (characters, facts, timeline) keeps later chapters consistent with earlier ones |
| Survives anything | Every step persists to disk - kill it mid-run and run resumes exactly where it stopped |
| Draws clear figures | The model authors a diagram spec; a layout engine (built-in, or D2 + ELK) places everything so labels never overflow or collide |
| Ships a file | Export to PDF · EPUB · DOCX · HTML · TXT · MD, with diagrams rendered as vector art |
Where to go next
Section titled “Where to go next”The TUI
The interactive shell: chat in plain English, slash commands, live run
dashboard, eleven themes (incl. the colourblind-safe highcontrast).
Tour the TUI →
Quality machinery
Thesis, a side-by-side draft judge, an insight gate, claim↔source checks, voice exemplars - how the ceiling is raised, not just the floor. How it works →
Commands
Every command and flag, from write to eval.
Reference →
Built different
Section titled “Built different”- Terminal or browser - the same pipeline runs in the interactive TUI or a local web dashboard (
writing-agent web,127.0.0.1only): Studio, a live SSE run view, per-project evals, artifacts, cost, and exports. - Cross-platform - Linux · macOS · Windows, CI-tested on Python 3.10–3.13.
- Zero required heavy deps - the core runs on stdlib + a handful of small libraries; deep-research fetching and embeddings are optional extras. The web dashboard is pure stdlib.
- Offline fake mode -
WRITINGAGENT_FAKE=1runs the entire pipeline with canned outputs: try the state machine, TUI, and exports without an API key. - MIT licensed - open source on GitHub, with every command, setting, and model route documented here.