How to schedule AI tasks in Claude Cowork and Codex

Executive overview

Most people interact with AI reactively. Claude Cowork and OpenAI Codex both support scheduled automations that run prompts on a recurring basis — hourly, daily, or weekly — without manual involvement.

The failure rate is high because people skip scoping, skip testing, and let output pile up inside the tool where nobody reads it.

Output must come to you — not the other way around.

Choosing between Claude Cowork and Codex

• Claude Cowork: accessed via the Claude desktop app under the Cowork tab → Scheduled
• Codex: accessed via the Codex desktop app under Automations
• Both are functionally equivalent; use whichever platform you already have
• Key difference: Cowork requires you to set "act without asking" manually; Codex enables autonomous action by default
• Both require the desktop app to remain open (not quit) and the computer to stay awake
• If the app was closed and a run was missed, both tools retroactively catch up when reopened

Setting up an automation

• Give it a name and (in Cowork) a short description — this is not the prompt
• Write the prompt: the most important field; be specific about every recurring step
• Choose a folder context (local folder) or run in chat (no folder)
• Set cadence: hourly, daily, weekdays, weekly, or manual
• Select model and reasoning level based on task complexity
• In Cowork, call skills with / and plugins; in Codex, use @ for apps and skills
• Set Cowork to "act without asking" — otherwise it pauses for permission on every action

Three questions before automating anything

1. Does this need to run on a clock?
2. Will someone act on the output?
3. Can the AI complete this with low risk — financially, legally, reputationally?

Only automate if the answer to all three is yes.

Scoping: keep automations to 1–4 steps

• Accuracy drops sharply as the number of steps increases
• A task with 5–20 steps has a high probability of failure
• One focused job per automation; resist the urge to bundle

Why automations fail

• Never runs: app was quit or computer went to sleep
• Can't reach the source: connector was de-authenticated, or the target folder was renamed or moved — fix by reconnecting the connector and restoring the folder path
• Output goes nowhere: the result sits inside Cowork or Codex and nobody checks it — fix by routing output to email drafts, Slack, Google Sheets, or another location you already monitor

Four high-value use cases

1. Monday briefing — AI reads calendar and inbox for the past and next seven days, drafts a weekly priorities summary to Gmail/Outlook drafts
2. Dashboard updater — AI monitors an input folder; when new data arrives, it updates an Excel sheet or dashboard, then moves the source data to an archive folder
3. Meeting prep — AI checks next-day meetings, attendees, inbox follow-ups, and past meeting transcripts; delivers a brief the evening before
4. CRM follow-up drafter — AI scans CRM weekly for silent prospects, drafts personalised follow-up emails to the drafts folder for human review before sending

Building automations correctly

• Run the task manually in a regular chat first — work through every step with the AI until the output is right
• Have the AI write a reusable prompt at the end of that chat that captures the full process
• Create the scheduled task and click "Run Now" — do not set a cadence yet
• Verify the output is correct
• Only then set the recurring cadence

Heartbeat and watchdog automations

Create these two meta-automations regardless of what else you build:

• Heartbeat: runs daily, writes the current time and date to a small log file — if it stops appearing, you know the app was closed and all other automations also stopped
• Watchdog: runs weekly, reviews all other scheduled tasks and writes a one-line status per task (ran successfully / failed / silent) — use the log to identify broken automations and fix them

Monthly audit

• Review every scheduled task and ask: is this still earning its spot?
• Delete anything you are no longer checking or acting on
• Automation accumulation is a silent cost — prune aggressively