Skip to content

RESEARCHCLAW_AGENTS.md

Latest commit

 

History

History
85 lines (64 loc) · 3.69 KB

RESEARCHCLAW_AGENTS.md

File metadata and controls

85 lines (64 loc) · 3.69 KB
Raw

AutoResearchClaw — Agent Configuration

Overview

ResearchClaw is an autonomous research pipeline that takes a research topic and produces a complete academic paper through 23 automated stages. This file defines how AI agents should bootstrap and interact with the system.

Agent Role: Research Orchestrator

You are an AI research assistant operating ResearchClaw. Your job is to:

  1. Understand the user's research interest — clarify the topic, scope, and constraints
  2. Configure the pipeline — set up config.yaml with appropriate LLM settings and experiment mode
  3. Execute the pipeline — run the 23-stage pipeline via CLI or Python API
  4. Monitor and intervene — handle gate stages (5, 9, 20), review intermediate outputs
  5. Deliver results — present the final paper, charts, and experiment data to the user

Quick Setup

# Install
pip install -e .

# Configure (copy and edit)
cp config.researchclaw.example.yaml config.yaml
# Set llm.base_url, llm.api_key, experiment.mode

# Run
researchclaw run --topic "Your topic" --auto-approve

Pipeline Stages (23 stages, 8 phases)

Phase Stages Description
A: Research Scoping 1-2 Define topic, decompose into sub-problems
B: Literature Discovery 3-6 Search strategy, collect papers, screen [GATE@5], extract knowledge
C: Knowledge Synthesis 7-8 Cluster topics, generate hypotheses
D: Experiment Design 9-11 Design experiments [GATE@9], generate code, plan resources
E: Experiment Execution 12-13 Run experiments, iterative refinement
F: Analysis & Decision 14-15 Analyze results, decide proceed/pivot/iterate
G: Paper Writing 16-19 Outline, draft, peer review, revision
H: Finalization 20-23 Quality gate [GATE@20], archive, export with charts, citation verification

Gate Stages

Three stages require approval (use --auto-approve for fully autonomous mode):

  • Stage 5 (Literature Screen): Validates collected literature quality
  • Stage 9 (Experiment Design): Validates experiment protocol before code generation
  • Stage 20 (Quality Gate): Validates overall paper quality before export

Experiment Modes

  • simulated: LLM generates synthetic results (fast, no code execution)
  • sandbox: Execute generated code locally (requires Python environment)
  • ssh_remote: Execute on remote GPU server (requires SSH configuration)

Key Files

File Purpose
config.yaml Pipeline configuration (LLM, experiment mode, etc.)
config.researchclaw.example.yaml Configuration template
researchclaw/cli.py CLI entry point
researchclaw/pipeline/executor.py Stage execution logic
researchclaw/pipeline/runner.py Pipeline orchestration
researchclaw/experiment/validator.py Code validation (AST, security, imports)
researchclaw/experiment/visualize.py Chart generation

Decision Guide

Situation Action
User provides a clear topic Run full pipeline with --auto-approve
User wants to review stages Run without --auto-approve, pause at gates
Pipeline fails at a stage Check error, fix config or retry from that stage with --from-stage
User wants iteration Use execute_iterative_pipeline() with max_iterations
Experiment code fails Validator auto-retries up to 3 times; if still failing, switch to simulated mode

Integration Platforms

ResearchClaw works with:

  • Claude Code: Load via .claude/skills/researchclaw/SKILL.md
  • OpenClaw: Read this AGENTS.md + README.md for bootstrapping
  • OpenCode: Compatible skill format in .claude/skills/
  • Standalone: Direct CLI or Python API usage