Summary
Before sending another instruction to Codex, first check whether the request changes work in progress and whether the result has already been saved or published. The timing can change how much of the work needs to be repeated.
- Use Steer to change what the current run is producing or how it proceeds.
- Use the Queue for work that should start after the current task ends.
- After you stop a run, first check what was saved to files and what changes reached external systems.

When to send an instruction
Steer sends a new instruction to the run in progress. Instructions sent to the Queue start a new run after the current one ends. Stop ends the current run. Check files saved or posts published before the stop in the system that holds them. OpenAI’s guide to Steer and Queue · How Codex runs and stops tasks
| Option | When to use it | Example instruction |
|---|---|---|
| Steer | The result in progress or the method used to produce it changes. | Before publishing, change the download link to “Original report.” Keep the attached PDF. |
| Queue | A separate feature or deliverable should be handled after the current task. | Once you verify the attendance calculation, design the Q&A board interface and permissions. |
| Stop | You need to know which changes took effect before running anything else. | Press Stop, then ask: “Before resuming, tell me which files were saved and which changes are still on the server.” |
- ChatGPT desktop app. Go to Settings → General → Follow-up behavior to choose the default. To send a single message differently, use the shortcut shown in Settings. You can edit or reorder queued messages above the composer.
- Codex CLI. During a run, press Enter to steer it or Tab to queue the message. OpenAI’s app and CLI instructions
Edits before publication and updates afterward
For one WordPress post, I changed the wording before publication and corrected a cost calculation afterward.

- Editing the wording before publication. While the publish screen was open, I asked Codex to change the download label from “Original paper” to “Original report.” Codex went back to the editor, changed the label, and published the post with the same PDF attached. It finished in one run.
- Correcting the cost calculation after publication. After the post went live, I found that it compared the entire monthly subscription fee with a portion of one week’s usage. In a new run, Codex calculated the share of the monthly fee equivalent to the same volume and updated the table and conclusion in the live post.
Before publication, the edit took place in the draft. Afterward, Codex had to reopen the live post and update the calculation as well. The edits differed in both timing and scope. The record shows which steps had to run again in each case. Published post with the corrected calculations
Changing the current run and queuing separate work
- Changing how issues are created in GitLab. While setting up the integration, I decided to use the API to create issues and use the browser only to issue the first token. Codex completed authentication setup in the same run. I saved the procedure in a skill so future issues follow the same steps.
- Changing the diagram library layout. While Codex was revising the rankings and detail views for 1,389 diagrams, I asked it to display more images at once. In the same run, it switched to six columns on wide screens and four on laptops. It also updated the detail view and keyboard navigation.
- A separate feature for the course support site. While Codex was correcting the attendance calculation, I asked for a new Q&A board. The request remained in the queue at the time. The board needed its own decisions about student and administrator viewing permissions and file attachments.
Late intervention and what remains after stopping
When the lookup method was specified
- First run. In about three minutes, Codex used the API to find one project, five running deployments, and four IP addresses matching the allocation records. It then inspected the admin interface. About 24 minutes after the run started, I instructed it to use only the API, and the run was stopped.
- Second run. On another PC, I specified a skill that uses only the API. Codex returned the results in about three minutes.
Both runs took about three minutes to obtain the necessary API data. The first went on to inspect the admin interface. The second ended after the API lookup. Choosing the route from the start could cut down on extra investigation. Because the devices and stopping points differed, the total time difference cannot isolate the effect of the skill or steering.
Search index state after stopping
While testing search in development, we removed the existing index. The run stopped before a new index was created.

- Work completed before the stop. The existing search index was removed.
- Results of the follow-up check. The database held 990 records, and the index had zero entries.
- Work remaining. We checked whether the stored records could be used to rebuild the index. Rebuilding it remained the next step.
After stopping, check both how far the run got and where its results were saved. In this case, the original data was still available, so the work could resume with rebuilding the index.
Changes to steering defaults and message delivery
The Codex app added steering on February 5, 2026. The February 8 app release changed the default for follow-up messages to Queue. February 5 app changelog · February 8 app changelog
An August 22 settings backup from the Mac used for these records showed Queue selected. On November 11, the same Mac showed Steer as the default and had a record that the previous setting had changed. Those records do not narrow down the exact date of the switch.
- October 8 desktop app. The app fixed a problem that left queued messages stuck during delivery and restored steering in secondary windows. Desktop app changelog
- October 15 Codex CLI. Follow-up messages could reach the current run immediately, even while it generated a response. Stopping a run also began stopping active Code Mode work and its tool calls. CLI 0.163.0 changelog
- November 2 iOS app. Without a saved user preference, follow-up messages went to the current run by default. iOS app changelog
With Steer as the default, users have fewer occasions to choose a delivery mode manually. Seeing the publish screen or knowing which tool was in use also helped me specify the timing and target of the edit, which may have helped Codex apply it in the same run.
Measuring the effect of the app change would require repeating the same work with the same follow-up instructions. The conversation records show when the instructions arrived and what work changed afterward. Instruction-following rates and changes in total work time require separate measurement.
Writing instructions during a run and measuring their effects
- Timing. Say when the change should happen, for example, “before publishing.”
- What to change. Specify the edit, for example, “change the download link text to ‘Original report.’”
- What to keep. State what should stay in place, for example, “keep the existing PDF.”
Save repeatable procedures, such as GitLab issue creation through the API, in a skill. To compare Steer and Queue, send the same correction at the same stage. Then record the time to reach a usable result, the time the user spent directly on the work, and the number of steps repeated.
Series overview · How the Way You Use Codex Affects the Outcome
