Context disappears
A new session does not know what the last session decided.
The problem
AI can do the work and still lose the thread.
Research files, decisions, and results often end up in different places. That makes an AI assistant repeat work or state a conclusion without showing how it was produced.
Files drift
Code, notes, and outputs spread across folders with no clear current copy.
Results lose their method
A number survives, but the exact code, data, and settings do not stay beside it.
Claims outrun evidence
A draft can sound certain even when no recorded experiment supports it.
How it works
Follow any claim back to what ran.
cleanResearch keeps an evidence chain: each conclusion points to a recorded experiment, and each experiment preserves the files needed to inspect it.
Claim
Insight
Experiment
Code + data + configuration
Configuration means the saved settings used for a run. “Frozen” means the exact executed files are preserved even if the live code changes later.
Six built-in skills
Ask in plain language.
A skill is a saved checklist that tells your AI how to handle one research task.
Welcome
Starts the project with a short interview and records your research context.
Help
Explains the workspace and gives you one direct next action.
Experiment
Writes the expectation first, runs the test, and preserves exactly what ran.
Audit
Checks every claim, result, figure, link, and experiment record for missing evidence.
Status
Shows the question, hypotheses, recent experiments, newest insight, and next step on one screen.
Tidy
Proposes safe file moves, asks once, and archives old work instead of deleting it.
A conversation with receipts
The record grows as you work.
Researcher
Does caffeine timing affect my assay yield?
AI
I’ll record this as experiment 002. Before I run it, I’ll write the expected effect and what result would count against it. Then I’ll save the exact command, inputs, settings, outputs, and executed code in experiments/002-caffeine-timing/.
Researcher
What evidence supports H1?
AI
H1 is supported only within the generated example data. The chain is hypotheses.md → insights.md → experiments/001-first-example/EXPERIMENT.md → analyze_steps.py, config.json, data/raw/example-steps.csv, and outputs/metrics.json. The recorded difference is 2,800 steps per day.
Start in 30 seconds
Open the folder with your AI coding assistant.
Paste this prompt. The assistant will read the project rules and ask one short question at a time.
Read AGENTS.md and follow its rules. Start the cleanResearch welcome interview. Ask one question at a time, use plain language, and do not change any research hypothesis without my confirmation.
中文简介
cleanResearch：面向 AI 的研究工作空间
cleanResearch 让 AI 助手持续理解你的项目，并为每个结果保留完整证据链：主张 → 研究结论 → 实验 → 当时使用的代码、数据与配置。
你可以直接用自然语言让 AI 建立项目、运行实验、检查证据、查看状态或整理文件。每次实验都会先记录预期，再运行分析，并保存实际执行的文件。
在 GitHub 上开始使用