Brooke Wright
Brooke Wright · @wright_mode
FREE CODEX GUIDE

Set up your first Codex automation

Put a recurring business job on a schedule. Get the prompts for transcript recaps and enquiry drafts, plus the checks that show it worked.

Copyable promptsBeginner friendlyBusiness examples
Something went wrong. Please try again.

No spam. Unsubscribe anytime.

5 copyable prompts for your first Codex automation Get the prompts → Join the Membership
Wright Mode — Free guide

Put Codex to work on a schedule

Tell Codex what to check, when to check it, and what to do with anything new.

Beginner friendlyCopyable promptsFolder and email examples

What's inside

There is a very boring business job you probably do more often than you need to. Checking for a new transcript. Scanning your inbox. Turning the same kind of enquiry into a draft.

Give Codex that job and a schedule. Start with one.

Pick a prompt. Replace the brackets. Paste it into Codex.

The first prompt is a fill-in template. The next two are complete business examples. Use folders inside the project you have opened, and name your time zone.

Your fill-in automation template

Use this for your own recurring job. Replace every [BRACKET] before pasting. The record location can be a file inside your project, such as processed-items.json.

Create a Codex automation called [NAME], attached to this chat. Every [INTERVAL, OR DAYS AND TIME] in [TIME ZONE], check [EXACT FOLDER OR CONNECTED SOURCE] for [WHAT COUNTS AS A MATCH]. For each item you have not already processed, [EXACT JOB]. Save [OUTPUT FORMAT] in [OUTPUT LOCATION]. Keep a record of successful items at [RECORD LOCATION]. Use a file hash or the email message ID to identify each item, so later runs skip it. Record success only after the output is saved. Include existing matching items on the first run. If there is nothing new, stay silent. Report completed work, errors or anything you need from me in this chat. Before scheduling it, test access to the source and output location. Read back the saved name, schedule, time zone, instructions and active status after creating it.

New transcript → client recap

Open a new test project in Codex. Replace the time zone, such as Australia/Perth. This prompt uses the folders incoming and recaps inside that project.

Create a Codex automation called "Client transcript recaps", attached to this chat and the project folder I have opened here. Every hour, in [YOUR TIME ZONE], check the incoming subfolder for complete .txt or .md transcript files. Include existing matching files on the first run. For each new file, read the full transcript. Create a recap with: what was discussed, decisions, actions with owners and due dates, and open questions. Use only what the transcript says. Write "not stated" for missing owners or dates. Save each recap in the recaps subfolder, using the source filename with -recap before .md. Keep processed-files.json in the project folder, recording the source file hash and output path only after saving successfully. Skip hashes already recorded. Keep source files in place. If nothing is new, stay silent. Report saved recaps or errors here. If a file is unreadable or still changing, leave it for a later check. Before scheduling, create incoming and recaps if needed. Test access to the folders and the ability to save a recap. After creating it, read back its name, hourly schedule, time zone and active status.

New enquiry → reply draft

Use your connected Gmail account and a narrow search. For example, label:client-enquiries, after you have created that label. Your service notes give Codex the facts it needs to draft a useful reply.

Create a Codex automation called "Client enquiry drafts", attached to this chat and the project folder I have opened here. Every weekday at 9 am in [YOUR TIME ZONE], check my connected [GMAIL ACCOUNT] for messages matching [GMAIL SEARCH OR LABEL], received since [START DATE]. For each message whose ID is not in processed-emails.json in this project, read the enquiry and the relevant thread. Write a reply draft using service-notes.md in this project, which contains my business facts and voice notes. Use only confirmed information from those notes. Flag anything missing; do not invent prices, availability or promises. Save each draft as a separate .md file in the drafts subfolder. Include the subject, email message ID and draft text. Do not send it or change the email. Record the message ID and output path only after saving successfully. If no matching unprocessed messages arrive, stay silent. Report saved drafts, errors or anything you need from me here. Before scheduling, confirm you can read this Gmail account and service-notes.md. Create drafts if needed and test saving a file there. After creating it, read back the name, schedule, time zone and active status.

The timing is simple

Scheduled check → new item found → job runs → result saved. An hourly check picks up a file or email on its next run.

1

Open your working folder

Sign in to Codex desktop and open a project folder. Use a folder where Codex can read the input and save the output.

2

Make the job work once

Ask Codex to try the job with one sample. Check the result before adding a schedule.

3

Paste your completed creation prompt

Give the job a name. Include the source, matching rule, schedule, time zone and output location.

4

Check the saved automation

Ask Codex to read back its details. Find it under Automations or Scheduled in your app and check that it is active.

5

Review the first scheduled result

Keep your computer awake and the app running for work that needs local files. Inspect the run and its saved output.

Return to this chat

Say “attached to this chat”. This suits a workflow you want to keep discussing and improving here.

Start a fresh run each time

Say “create a standalone automation”. Put all essential instructions in the saved prompt.

OpenAI now calls these Scheduled tasks in its documentation; your app may show Automations. Official setup and task types →


Use a small folder setup you can recognise later. Keep the output separate so a recap never gets mistaken for another transcript.

📥

incoming

New .txt or .md transcripts land here.

📄

recaps

Finished recaps are saved here.

✅

processed-files.json

A record of successful files, so the next run can skip them.

Try this with a fictional transcript

Save the text below as sample-call.txt in a new test incoming folder.

Client: We need the workshop outline by Friday. Brooke: I will send the outline on Thursday. Client: I will confirm attendee numbers tomorrow. Brooke: We still need to choose the venue.

The recap should keep both actions, leave relative dates as stated, and flag the venue as an open question.

Tell it what to do with existing files

The supplied folder prompt includes existing matching files on its first run. If you only want future arrivals, ask Codex to record the existing files as a starting baseline before activating it.

Done when: one new transcript produces one recap, the source stays in place, and another check skips it.


“Check my email” leaves too much room for guessing. Name the account and tell Codex which messages count.

1

Connect the inbox you actually use

Connect Gmail in your app. Ask Codex to show one matching subject line before scheduling anything; this proves the connection can read it.

2

Narrow the match

Choose an existing label, sender or subject. In the supplied prompt, replace the search, start date and account placeholders.

3

Give it your business facts

Save your business facts and voice notes as service-notes.md in the project. Include the information you would use when writing the reply yourself.

4

Save the draft where you can review it

The supplied prompt writes drafts into your project folder. It uses the email message ID to avoid repeating successful items.

Check both kinds of access

Reading Gmail and saving a file are separate actions. Test both in the chat and inspect the first scheduled run. If an app connection or folder is unavailable, the workflow needs that fixed first.

This guide uses scheduled inbox checks in Codex desktop. OpenAI also supports eligible email event triggers on web and mobile. Check event-trigger availability →

Done when: one matching enquiry produces a reviewable draft with the original message ID, and the next check skips it.


An automation appearing in the list proves it was created. Check what it actually did.

Test the job and check for repeats

Use this in the same chat after creating the automation. Review a scheduled run too; a successful manual check alone does not prove the schedule fired.

Run one manual check using the saved instructions for [AUTOMATION NAME]. Do not create another automation. Show me what you found, the exact output location, and the successful-item record. Then repeat the same check with no new input and confirm the item is skipped rather than processed twice. If anything failed, explain the cause and what access or information is missing. A manual check is a workflow test; also show me the saved schedule and where to review the next scheduled run.

Your “it works” checklist

  • The saved name, schedule and time zone match your request.
  • A scheduled run appears at the expected time.
  • The output exists where you asked for it.
  • Its content matches the source.
  • The successful-item record includes that input.
  • A repeat check produces no duplicate.

Change the schedule

Name the existing automation so Codex updates it. You can also say “Pause [AUTOMATION NAME]”, “Resume [AUTOMATION NAME]”, or “Show me its saved instructions and status.”

Update the existing Codex automation [AUTOMATION NAME] to check [SOURCE] every [NEW SCHEDULE] in [TIME ZONE]. Keep its other instructions and successful-item record. Read back the saved schedule, time zone and status. Do not create a second automation.

No result?

  • Check whether the automation is active.
  • For local files, check the computer and app were running.
  • Check the folder path or email search matched something.
  • Read the run result for missing access or an error.

The same item keeps appearing?

  • Check the successful-item record was saved.
  • Use a file hash or message ID as the identifier.
  • Check a second automation is not doing the same job.
  • Keep outputs outside the incoming folder.

Checked against OpenAI’s official scheduled-task documentation on 01/10/2026. The prompts and folder examples are Wright Mode’s practical templates. The interface and available connections can vary by account.


Build your next workflow with me

Bring the business job you want to automate to Wright Mode Membership.