A practical AI-assisted research workflow for using Codex to read papers, manage ideas, edit code, run experiments, and keep reproducible research logs.
面向科研新手、论文复现、模型改进和实验管理的 Codex 科研工作流模板。它不是一个自动写论文工具，而是一套让你和 AI coding agent 协作时不乱改、不丢结论、可复现、可回溯的 Markdown 工作台。
很多人开始用 Codex、Claude Code 或其他 AI coding agent 做科研后，会很快遇到这些问题：
- 读了很多论文，但创新点没有进入可执行列表；
- agent 改了代码，但不知道改动动机和实验假设是什么；
- 小测试、正式实验、消融实验混在一起，最后无法解释结果；
- baseline 被覆盖，失败实验没有记录，复盘时找不到证据；
- 跨设备、跨分支、跨协作者同步时，研究状态断层。
这个 starter kit 的目标是把科研过程拆成一个清晰闭环：
problem
-> paper evidence
-> idea pool
-> scoped implementation
-> smoke test
-> formal experiment
-> result review
-> keep / revise / drop
- 正在复现或改进深度学习论文的学生；
- 想用 Codex 辅助科研代码修改的人；
- 需要管理 paper notes、idea pool、experiment logs 的研究者；
- 和同学、导师、远程服务器协作时，需要同步研究状态的人；
- 想把失败实验也记录下来，避免重复踩坑的人。
| File | Purpose |
|---|---|
WORKFLOW.md |
从问题定义到实验复盘的完整科研闭环 |
AGENTS.md |
给 Codex / AI agent 看的协作规则 |
templates/paper_note_template.md |
论文阅读和代码对齐模板 |
templates/innovation_pool_template.md |
创新点候选池和消融设计模板 |
templates/experiment_log_template.md |
正式实验记录模板 |
templates/model_change_log_template.md |
模型/代码修改记录模板 |
templates/weekly_plan_template.md |
周计划和周末复盘模板 |
templates/sync_log_template.md |
跨设备、跨分支、跨人员同步模板 |
notes/README.md |
真实研究记录的存放说明 |
- Clone or copy this repository into your research project.
- Read
WORKFLOW.mdbefore asking Codex to modify code. - Copy a template from
templates/intonotes/. - Rename it by date or topic, for example:
notes/paper_notes/2026-06-05-transformer-baseline.md
notes/experiment_logs/2026-06-05-baseline-smoke-test.md
- When you ask Codex to work, point it to the relevant notes:
Please read AGENTS.md, WORKFLOW.md, and notes/innovation_pool/2026-06-05-idea-pool.md first.
Implement only I-002 and record the model change.
You are helping me with a research project.
Before editing code, read:
- AGENTS.md
- WORKFLOW.md
- the related paper note
- the current experiment log
Rules:
- Keep the baseline unchanged.
- Change only one core variable at a time.
- Run a small smoke test before proposing formal experiments.
- Update the model change log and experiment log after the change.
- Do not upload datasets, weights, tokens, private paths, or temporary outputs.
my-research-project/
├─ AGENTS.md
├─ WORKFLOW.md
├─ templates/
│ ├─ paper_note_template.md
│ ├─ innovation_pool_template.md
│ ├─ experiment_log_template.md
│ ├─ model_change_log_template.md
│ ├─ weekly_plan_template.md
│ └─ sync_log_template.md
├─ notes/
│ ├─ paper_notes/
│ ├─ innovation_pool/
│ ├─ experiment_logs/
│ ├─ model_change_logs/
│ ├─ weekly_plans/
│ └─ sync_logs/
└─ your_actual_code/
- Read papers and code before inventing new modules.
- Keep every idea connected to a source, motivation, code location, and validation plan.
- Preserve the baseline and implement new methods as independent variants when possible.
- Change one core variable at a time so ablation results are meaningful.
- Treat smoke tests as code checks, not paper-level conclusions.
- Record failed experiments because they protect you from repeating dead ends.
- Never publish datasets, model weights, private repositories, tokens, or unpublished research details.
These projects inspired the public positioning and naming:
| Project | Stars checked on 2026-06-05 | Notes |
|---|---|---|
| Alibaba-NLP/DeepResearch | 19k+ | Popular deep research agent category |
| dzhng/deep-research | 19k+ | Simple deep research agent implementation |
| Orchestra-Research/AI-Research-SKILLs | 9k+ | AI research skills for coding agents |
| jhfnetboy/DSR-Research-Flow-Template | 70+ | Close workflow-template reference |
| KaleabTessera/Research-Paper-Reading-Template | 70+ | Paper note template reference |
This repository is intentionally different from deep research agents: it focuses on human-guided, reproducible research operations with Codex.
这是一个“用 Codex 做科研”的工作流模板，适合论文复现、模型改进、实验记录和阶段同步。你可以把它复制到自己的科研项目里，让 Codex 每次改代码前先读 AGENTS.md 和 WORKFLOW.md，并在改动后同步更新实验记录与代码修改记录。
MIT License. You can copy, modify, and adapt it for your own research projects.