
OpenClaw Creator Barely Lets Claude Write Code: Claude Plans, Codex Executes
Vũ Khôi
Tác giả tại Wotech Electronics Blog

OpenClaw Creator Barely Lets Claude Write Code: Claude Plans, Codex Executes
Peter Steinberger, the creator of OpenClaw, almost never uses Claude to write code directly. Instead, he has a more efficient workflow: Claude plans, Codex executes. That is the content of a post by the Học lập trình cho người mới bắt đầu page (19/11/2026). The reason is simple, per the post: Claude's tokens are limited and expensive. So in Peter's system, Claude coordinates the work, while Codex writes the code. All workflow details, figures, and views in this article follow the post and have not been independently verified by Wotech.
Quick answer: According to the post, Peter Steinberger's workflow has 4 steps: (1) Claude analyzes and designs the change, then creates a detailed order covering the goal, the files to edit, the parts that must absolutely not be touched, and the tests proving the change works; (2) each order goes to a brand-new Codex session to execute independently — last month, Peter said he ran up to 27 Codex sessions at once; (3) Codex takes on nearly all technical work — writing code, fixing CI errors, rebasing, and resolving merge conflicts — then a second Codex session reviews the code just created; (4) finally Claude checks the result, evaluates the evidence from the tests, and decides whether to merge. One rule in the system: Claude directly edits code only when the change is under 20 lines; larger work all goes to Codex. The whole workflow is packaged as a Skill named codex-first.
OpenClaw is the open-source project created by Peter Steinberger — per the post, he is the "creator of OpenClaw".
How does the 4-step workflow work?
According to the post, the workflow works as follows:
- Claude plans. Claude analyzes and designs the change. Then Claude creates a detailed order covering the goal, the files to edit, the parts that must absolutely not be touched, and the tests proving the change works.
- Codex executes, in parallel. Each order goes to a brand-new Codex session to execute independently. Last month, Peter said he ran up to 27 Codex sessions at once.
- Codex reviews Codex. Codex takes on nearly all technical work: writing code, fixing CI errors, rebasing, and resolving merge conflicts. Then a second Codex session reviews the code just created.
- Claude decides on the merge. Finally, Claude checks the result, evaluates the evidence from the tests, and decides whether to merge.
Who does what in this system?
According to the infographic in the post, the roles are clearly split:
| Role | Claude | Codex |
|---|---|---|
| Thinks | Designs the change and the API; writes orders for every task; checks the proof; decides what gets merged | — |
| Types | — | Writes the code from the order; fixes bugs, tests, refactors, migrates; fixes CI, rebases, resolves conflicts; reviews the diff in a separate session |
One interesting rule in the system, per the post: Claude directly edits code only when the change is under 20 lines. Larger work all goes to Codex.
Apply now: try the codex-first model
According to the post, the whole workflow is packaged as a Skill named codex-first. To try this working model, you only need to install the Codex CLI, then copy the Skill folder into ~/.claude/skills. The repo and detailed guide, per the post: https://github.com/steipete/agent-scripts/blob/2f82700ba9f3e7e93d3583e1990992584561eadc/skills/codex-first/SKILL.md
A thought-provoking approach, per the post: instead of making one AI do everything, let each model do exactly the role it does best. Claude is the architect and reviewer, Codex is the executing engineer.
Frequently asked questions
1. What are the steps of Peter Steinberger's codex-first workflow? According to the post: (1) Claude plans and writes a detailed order; (2) each order goes to a fresh Codex session to execute independently, runnable in parallel; (3) another Codex session reviews the code; (4) Claude checks the test evidence and decides whether to merge.
2. Why doesn't Claude write code directly? According to the post, Claude's tokens are limited and expensive, so Claude's expensive tokens go to thinking, while Codex does the typing.
3. How many Codex sessions did Peter Steinberger run at once? According to the post, last month he said he ran up to 27 Codex sessions simultaneously.
4. When does Claude directly edit code? According to the post, only when the change is under 20 lines; larger work all goes to Codex.
5. How can I try this model?
According to the post: install the Codex CLI, copy the codex-first Skill folder into ~/.claude/skills; the repo and detailed guide are at github.com/steipete/agent-scripts (skills/codex-first folder).
Source: post by the Học lập trình cho người mới bắt đầu page (19/11/2026) about Peter Steinberger's codex-first workflow, creator of OpenClaw. All workflow details, figures, and views in this article follow the post and have not been independently verified by Wotech.
Bài viết liên quan

20 tháng 11, 2026 · 10 phut doc
Java Dev, 32, Laid Off 23 Days Before His Wedding: 187 Applications, 0 Offers, and a 404 Not Found
A 32-year-old Java developer in Hangzhou was laid off exactly 23 days before his wedding — and the ending is what leaves you sitting in silence. The story was shared by the account Stanley (@Stanleys…

19 tháng 11, 2026 · 7 phut doc
Gemini 4 Neon Beats Opus 5.5 on the Artificial Analysis Board: Cheaper, Less Hallucination — But Not Everywhere
Neon launched this morning, replacing Argon — and according to a post by Đỗ Anh Việt in the Nghiện Claude group (19/11/2026), Gemini 4 Neon beats Claude Opus 5.5 on most fronts on the Artificial Anal…

19 tháng 11, 2026 · 7 phut doc
First Director Criminally Tried in Vietnam for Using Unlicensed Software in His Own Company
No selling pirated software, no distributing cracked software — only using software to run his own company's operations. So why was the director still brought to criminal trial? On 17/11/2026, the Pe…