Schedule AI Prompts in Claude Co-Work or Codex Automations

Comparing Claude Co-Work Scheduled Tasks and Codex Automations

Claude Co-Work scheduled tasks and Codex automations both execute prompts at set intervals such as hourly, daily, weekday, or weekly. Users access them via desktop apps: in Claude, select Co-Work then Scheduled; in Codex, select Automations. Both require the computer to stay awake and the app running (minimized, not quit); Claude shows a 'keep awake' note, while Codex runs in background by default. Missed runs self-heal upon restart.

Creation interfaces are similar. Name the task, add a description in Claude (short purpose summary), and write the core prompt detailing the recurring action. Select a dedicated folder (e.g., on desktop) for inputs/outputs; Claude offers folder or general, Codex offers local (folder) or chat. Choose models like Claude Opus or Codex 5.5. Set reasoning to low/medium/high/extra high in Claude based on complexity. Crucially, enable 'act without asking' in Claude (or default in Codex) for autonomous tool use; otherwise, it prompts for permission each run, defeating automation.

Prompts invoke skills/plugins: in Claude, use '/' for skills/plugins (plugins group skills); in Codex, use '@' for apps/skills/plugins (color-coded: blue for skills, bluish-purple for plugins, red for apps). Codex feels more user-friendly for precise calls.

Criteria for Automatable Tasks and Common Failures

Automate only if three conditions hold: (1) the task needs a fixed clock (hourly/weekly without involvement); (2) the output triggers action (personal or system); (3) it runs autonomously with low financial/legal/reputational risk, connecting reliably to sources like email/calendar.

Limit tasks to 1-4 steps; more reduces accuracy (chart shows sharp drop beyond 4). Failures include: (1) never runs (computer sleeps/app quits—fix by preventing sleep/quit); (2) misses sources (renamed/moved folders or failed connectors—reconnect/reposition); (3) output stays in-app (unseen—route to email drafts, Slack DMs, Google Sheets, or checked folders so it reaches you).

Starter Use Cases

Four simple cases: (1) Monday briefing reads calendar/inbox for past 7/next 7 days, identifies tasks, drafts summary in Gmail/Outlook drafts. (2) Dashboard/Excel updates: place sheet and input folder (with new data) plus archive folder in dedicated dir; AI processes inputs, updates sheet, moves inputs to archive (daily/weekly). (3) Evening meeting prep scans tomorrow's calendar for attendees/follow-ups from inbox/transcripts/prior commitments, sends note. (4) CRM follow-ups (weekly/monthly) flags silent prospects, drafts personalized notes to inbox drafts.

Building Process and Monitoring

Build iteratively: (1) Prototype in chat (Claude/ChatGPT) until output satisfies, including connectors/tools. (2) Have AI generate a generalized prompt encapsulating the process. (3) Paste into new scheduled task/automation, set to manual, run and verify. (4) Then activate cadence.

Add two meta-tasks: Heartbeat (daily) logs timestamp to a file—check for runs. Watchdog (weekly) scans all tasks' outputs, notes issues to a file. Outputs must push to you (e.g., drafts/Slack/Sheets).