Open-FARS Template

Open-FARS (Fully Automated Research System) template repository. Scaffold a new AI-driven research project with a single command.

What is Open-FARS?

Open-FARS is a 7-stage automated research pipeline driven by Claude Code:

Stage	Agent	What it does
S1	open-fars-survey	Systematic literature survey via Semantic Scholar API
S2	open-fars-ideation	Generate and score research ideas from survey gaps
S3	open-fars-plan	Design experiments: RQs, hypotheses, baselines, task graph
S4	open-fars-assets	Implement method, baselines, evaluation code
S5	open-fars-experiment	Execute experiments, statistical analysis, figures
S6	open-fars-writing	Write complete LaTeX paper
S7	open-fars-revision	Simulated peer review and revision

The pipeline runs autonomously with quality gates (judge agent) at each stage, email notifications at decision points, and automatic escalation when issues arise.

Quick Start

Install

git clone https://github.com/hzt/open-fars-template.git
cd open-fars-template
./bin/fars install    # Symlink to /usr/local/bin

Create a new project

fars init ~/research/my-project

The interactive wizard will ask for:

• Research direction (slug, e.g. llm-alignment)
• Description (one line)
• Target venue (e.g. NeurIPS 2026)
• Language (en or zh)
• Seed queries (comma-separated search terms)
• Compute resources (describe your GPU setup)
• Email (optional, for notification skill)

Start the pipeline

cd ~/research/my-project
# Open Claude Code and say:
# "Start the Open-FARS pipeline"

Project Structure (after fars init)

your-project/
├── CLAUDE.md                    # Points to AGENTS.md
├── AGENTS.md                    # Pipeline orchestration protocol
├── .gitignore
├── .claude/
│   ├── SPEC.md                  # Engineering specification (single source of truth)
│   ├── agents/                  # 8 subagent definitions
│   │   ├── open-fars-survey.md
│   │   ├── open-fars-ideation.md
│   │   ├── open-fars-plan.md
│   │   ├── open-fars-assets.md
│   │   ├── open-fars-experiment.md
│   │   ├── open-fars-writing.md
│   │   ├── open-fars-revision.md
│   │   └── open-fars-judge.md
│   └── skills/                  # 4 user-invocable skills
│       ├── catchup/SKILL.md     # /catchup - research onboarding doc
│       ├── review/SKILL.md      # /review - trigger quality review
│       ├── status/SKILL.md      # /status - project progress report
│       └── email-notify/        # email notification system
│           ├── SKILL.md
│           ├── send-email.py
│           └── config.json.example
└── .open-fars/
    ├── config.yaml              # Your project configuration
    └── meta/
        └── registry.yaml        # Pipeline state tracking

Skills (slash commands in Claude Code)

Command	Description
/review	Trigger quality review (auto-detects what to review)
/status	Generate project progress report
/catchup	Generate research onboarding document
/email-notify	Send structured notification email

Configuration

Edit .open-fars/config.yaml after scaffolding to fine-tune:

• Survey parameters (paper counts, citation depth)
• Ideation constraints (idea count, scoring weights)
• Experiment limits (max experiments, ablations)
• Writing settings (template, page limit)
• Revision targets (simulated reviewers, target score)
• Compute resources (GPU description)
• Orchestration (auto-review, escalation thresholds)

Full schema documented in .claude/SPEC.md.

Requirements

• Claude Code CLI
• Git
• Bash 4+