6 Practical Tips for Using Codex Effectively in Research Projects

After experimenting with Codex on research‑oriented coding work such as paper reproduction, exploring high‑quality repositories, and refining course project features, the author proposes a concise workflow.

1. Read the context before writing code

For tasks like reproducing papers or unfamiliar repos, begin by asking Codex to examine the project and report its structure, entry points, commands, relevant files, and uncertainties. Example prompt:

先不要修改代码。请先阅读这个项目，告诉我：
1. 项目结构是什么
2. 主要入口在哪里
3. 训练/测试/运行命令分别是什么
4. 哪些文件最可能和当前任务有关
5. 你不确定的地方有哪些

This upfront reading prevents the model from acting on wrong assumptions.

2. Use AGENTS.md for long‑term rules, not task descriptions

AGENTS.md should contain persistent guidelines such as:

## 工作规则
- 默认先阅读相关代码和文档，再开始修改。
- 修改前先说明影响范围。
- 只做和当前任务相关的最小改动。
- 提交前必须运行最小验证。
- 说明用中文，代码、命令、文件名保持英文。
- 不要修改无关文件。

For paper‑reproduction projects, add an "实验原则" section emphasizing hypothesis‑driven experiments and avoiding low‑value ablations.

3. Always plan complex tasks first

When a task spans multiple files, ask Codex to produce a plan that lists the problem, affected modules, a 3‑7 step breakdown, validation steps, and potential error points before any code is written. Example prompt:

先不要写代码。请进入计划模式，帮我设计实现方案：
1. 明确这次要解决的问题
2. 列出会影响哪些模块
3. 拆成 3-7 个步骤
4. 每一步怎么验证
5. 哪些地方最容易出错
等我确认后再开始实现

This reveals misunderstandings early and makes the plan itself valuable.

4. Require Codex to verify facts before acting

When reproducing a paper or analyzing a repo, first ask for a summary of the core idea, key formulas/modules, consistency between implementation and paper, a minimal reproduction path, and potential risk points. Example prompt:

请先阅读论文和官方 repo，不要开始写代码。
先输出：
1. 论文核心 idea
2. 关键公式/模块
3. 官方实现和论文描述是否一致
4. 复现最小路径
5. 可能复现不出来的风险点

This ensures the model aligns with the source material before generating code.

5. Start a new session after each completed task

Long sessions accumulate stale assumptions. After finishing a task, summarize the changes, current state, and next steps in a single sentence, then start a new conversation with that summary before proceeding.

请用一句话总结刚完成的任务、修改过的文件、当前状态和下一步注意事项。
然后 /new， 在新会话里贴这个总结，再开始下一个任务。

6. Never trust Codex’s “done” claim without explicit verification

At the end of each task, request a detailed report of what was changed, why, which validations ran, what remains unverified, and what still needs human confirmation.

1. 修改了哪些文件
2. 每个文件为什么改
3. 运行了哪些验证
4. 还有哪些没验证
5. 哪些地方需要我人工确认

The author concludes that stability comes not from a magical prompt but from clear context, enduring rules, and concrete acceptance criteria.