# auto_writer

**Repository Path**: ming616/auto_writer

## Basic Information

- **Project Name**: auto_writer
- **Description**: agentic paper writer
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-09
- **Last Updated**: 2026-06-09

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Auto Writer

An AI-powered academic paper writing system based on **Agentic Team** architecture. Multiple specialized AI agents collaborate to plan, research, write, review, and edit academic papers, outputting multi-file LaTeX projects.

## Architecture

```
Topic ──► Planner ──► Outline ──► Researcher ──► References
                    ┌───── Blueprint ─────┐
                    │  (per-section spec)  │
                    ▼                     ▼
             SectionWriterAgent × 8 ──► WriterTeam ──► Draft
                    │                          ▲
                    │   Shared Context:         │
                    │   • Previous sections     │
                    │   • References            │
                    │   • Blueprint             │
                    ▼                          │
              Reviewer ──── feedback ───► Editor ──► .tex files
```

### Agent Roles

| Agent | Responsibility |
|-------|---------------|
| **Planner** | Creates `PaperBlueprint` (section specs, key points, questions) from topic |
| **Researcher** | Gathers references & related literature for each section |
| **WriterTeam** | Orchestrates 8 section-level writers with shared context |
| ├ `SectionWriterAgent` x8 | Each writes one section using templates + blueprint |
| **Reviewer** | Reviews content, runs harness validation, suggests revisions |
| **Editor** | Runs harness validation, exports multi-file LaTeX project |

### Section Writers (共享上下文协作)

Each section has its own dedicated `SectionWriterAgent`, running sequentially so later sections can reference earlier ones:

| Section | Key | File |
|---------|-----|------|
| Abstract | `abstract` | `0_abstract.tex` |
| Introduction | `introduction` | `1_introduction.tex` |
| Related Works | `related_works` | `2_related_works.tex` |
| Formulations | `formulations` | `3_formulations.tex` |
| Proposed Method | `proposed` | `4_proposed.tex` |
| Experiments | `experiments` | `5_experiments.tex` |
| Discussions | `discussions` | `6_discussions.tex` |
| Conclusions | `conclusions` | `7_conclusions.tex` |

## Module Structure

```
src/auto_writer/
├── core/               # Agent framework
│   ├── agent.py        # Agent base class + AgentContext (hooks, run)
│   ├── message.py      # Message, MessageRole
│   ├── task.py         # Task, TaskStatus
│   └── team.py         # Team + PaperWritingTeam (workflow, review loop)
├── agents/             # Agent implementations
│   ├── planner.py      # PlannerAgent → creates PaperBlueprint
│   ├── researcher.py   # ResearcherAgent → gathers references
│   ├── writer.py       # SectionWriterAgent × 8 + WriterTeam orchestrator
│   ├── reviewer.py     # ReviewerAgent → harness validation + feedback
│   └── editor.py       # EditorAgent → validation + multi-file export
├── models/
│   └── paper.py        # Paper, PaperSection, Reference, Author, SECTION_DEFS
├── templates/          # LaTeX skeleton templates
│   ├── base.py         # Template, SectionTemplate, PaperTemplate
│   └── sections.py     # Default templates for all 8 sections
├── prompts/            # LLM prompt management
│   ├── base.py         # Prompt, SystemPrompt, TaskPrompt, PromptManager
│   └── agents.py       # Default prompts per agent role
├── harness/            # Validation & constraint framework
│   ├── base.py         # Constraint, Violation, Severity, Harness
│   └── rules.py        # HasAllSections, SectionContentCheck, etc.
├── blueprint/          # Content plan specification
│   ├── base.py         # PaperBlueprint, SectionBlueprint, SubsectionBlueprint
│   └── defaults.py     # Default blueprints for all 8 sections
├── tools/
│   └── formatting.py   # LaTeX export (main.tex, section files, .bib)
├── config.py           # Project-wide configuration
└── __init__.py         # Public API exports
```

## Key Design Decisions

- **Blueprint ≠ Template ≠ Prompt ≠ Harness** — Each concern is separated: blueprint defines *what content*, template defines *LaTeX structure*, prompt defines *LLM instructions*, harness defines *validation rules*.
- **Per-section `SectionWriterAgent`** — Each section gets its own agent instance with shared context, enabling cross-referencing and collaborative writing.
- **`WriterTeam` orchestrator** — Manages 8 section writers sequentially, propagating `shared_state` for context sharing.
- **Template system** — Uses `{{placeholder}}` syntax (not Python f-strings) to avoid LaTeX escaping issues.
- **Multi-file LaTeX output** — Generates `main.tex`, 8 chapter `.tex` files, and `references.bib`.
- **Review loop** — Supports multiple revision cycles with writer → reviewer → editor.
- **pydantic v2** — All data models use pydantic v2 for validation.

## Quick Start

```bash
pip install -e ".[dev]"
```

```python
import asyncio
from auto_writer import PaperWritingTeam

async def main():
    team = PaperWritingTeam()
    result = await team.run(topic="Transformer architectures in NLP", max_revisions=1)
    print(f"Output: {result['output_dir']}")

asyncio.run(main())
```

## Development

```bash
pip install -e ".[dev]"
pytest tests/ -v
```

## License

MIT