Putting Claude on a schedule
Recurring agents that do real work while you are away from the keyboard. This playbook covers how to set up scheduled tasks in Claude Code for three patterns I use and recommend, a morning digest, a weekly SEO check, and a content-draft generator, plus the rule I use for when Make.com is the better tool. No theoretical use-case lists. Just the setup, the guardrails, and the failure modes.
## the process
- 1
Pick work that actually earns a schedule
A task belongs on a schedule when three things are true. It recurs on a clock, the prompt is identical every time, and the output is something you can review later instead of something you need mid-task. Inbox triage, weekly checks, and first drafts all qualify. One-off research and anything you would correct in real time do not. If you cannot describe the job in two sentences, it is not ready to run unattended.
- 2
Write the prompt like a brief, then test it live
Write the prompt the way you would brief a new assistant on their first day. One job, a defined output format, and a defined destination for that output. Then run it manually in a normal Claude Code session at least twice before scheduling it, because a scheduled run cannot stop and ask you what you meant. Every ambiguity you leave in the prompt becomes a decision the agent makes without you.
- 3
Wire up tool access before you schedule anything
If the agent needs your calendar, inbox, or an SEO tool, connect those MCP connectors first and confirm they work in an interactive session. A scheduled run cannot pause for an OAuth screen or a permission prompt, it just fails or skips the step. Grant the narrowest access that does the job. An inbox-triage agent needs to read mail and create drafts, it does not need send permission, so do not give it send permission.
- 4
Create the schedule in Claude Code
In Claude Code, describe the cadence in plain English, for example run this prompt every weekday at 7am, and it creates a scheduled task. Under the hood it is a cron schedule attached to your prompt, and the task runs as its own agent session in the cloud, so your machine does not need to be on. You can ask Claude Code to list your scheduled tasks, change a cadence, or delete one the same conversational way. Start with one schedule, not five.
- 5
Pattern one, the morning digest
Schedule a weekday early-morning run. The prompt pulls today's calendar events, scans email that arrived since the last run, and sorts messages into three buckets, needs a reply, worth knowing, and ignorable. For the needs-a-reply bucket, it writes responses and saves them as drafts in the mailbox, never sends. The output is one short digest you read with coffee, today's schedule, what needs you, and which drafts are waiting for approval.
- 6
Pattern two, the weekly SEO check
Schedule a Monday morning run. The prompt pulls current rankings and site-health data from your SEO tool, compares against the snapshot it saved last week, and flags only what moved, pages that dropped, errors that appeared, queries that grew. Cap the output at three recommended actions. The point of this agent is a weekly nudge that takes two minutes to read, not a report nobody opens.
- 7
Pattern three, the content-draft generator
Keep a topic backlog in a plain file or a board the agent can read. On a weekly schedule, the agent picks the top topic, writes a first draft following a voice-and-structure brief you wrote once, and saves it to a drafts folder. It never publishes. The win is not finished content, it is that you start every editing session at sixty percent instead of at a blank page.
- 8
Know when Make.com is the right tool instead
My rule is judgment versus plumbing. If the job is read, summarize, draft, or decide, that is Claude. If the job is when X happens copy the data to Y, with no judgment involved, that is Make.com, it is cheaper, instant on triggers instead of waiting for a clock, and it never hallucinates a field mapping. The two also combine well, Make catches events and files them somewhere Claude reads on its schedule, or Claude leaves an approved draft where a Make scenario picks it up and ships it.
- 9
Babysit it for two weeks before you trust it
For the first two weeks, read every output the schedule produces and note what you had to fix. Tighten the prompt after each miss, most failures trace back to an instruction you thought was obvious. Set a calendar reminder to audit your schedules monthly, and apply the dead-agent rule, if you have stopped reading an agent's output, delete the schedule. Unread automation is not productivity, it is spend.
## hard-won tips
- !Drafts, never sends. A scheduled agent should produce things you approve, not actions that already happened while you slept.
- !One job per agent. A digest agent that also does SEO checks and writes posts fails at all three in ways that are hard to debug.
- !Scheduled runs consume your usage while you are not watching. Check what each schedule costs you in the first week and cut anything where the review time exceeds the time saved.
- !Pin the output format in the prompt, with section names and a length cap. Unattended runs drift toward long and vague unless you fence them in.
- !Never test a new prompt by scheduling it. Run it interactively first, schedules are for prompts that already work.
$ follow --the-build
Watch it happen, don't take my word for it
Every build on this site gets documented as it happens — the prompts, the dead ends, the results. No course at the end of this funnel. There is no funnel.
follow on x →