Claude Code Routines
Set the work up once, then let it happen without you.
Kevin Lenida
A routine is Claude doing your work while you are not there.
The same task, two ways
You do it
- You have to remember
- You retype the same prompt
- It happens when you have time
- Skipped on your busiest days
A routine does it
- It remembers
- The prompt is already written
- It happens at eight, every day
- Busiest days are when it helps most
Four parts, one machine
Three ways a routine can start
- On a schedule — hourly, daily, weekdays, weekly, or once at a set time.
- By an API call — your own tools fire it the moment something happens.
- On a GitHub event — a pull request opens, a release ships.
- One routine can use all three at once.
- Most people only ever need the first one.
It runs on Anthropic's computers, so your laptop can be closed.
Two kinds. Only one runs without you.
Cloud routine
- Runs on Anthropic's computers
- Works with your laptop closed
- Made on the web, in the terminal, or in the desktop app
Local routine
- Runs on your own machine
- Only while that machine is awake
- Desktop app only — but it can reach your files
What you need
- A paid Claude plan — Pro, Max, Team, or Enterprise.
- Claude Code on the web enabled — routines live in your account.
- Connectors — Gmail, Calendar, Slack, Drive, whatever it must reach.
- A code repository — only if you want one. Email routines need none.
- Research preview — expect small changes, as of July 2026.
claude.ai/code/routines · claude.ai/customize/connectors · code.claude.com/docs/en/routines
Build one
A routine that triages your inbox every morning.
Start from blank, or from a template
Email triage is a built-in template — close to what we are about to build by hand.
Five things to fill in
- Name — the only field you are actually needed to fill.
- Instructions — the prompt it runs, unchanged, every single time.
- A repository — optional. Skip it for anything that is not code.
- A trigger — when it should start.
- Connectors — what it can touch.
You never write a cron string
Pick a preset, pick a time. That is the entire schedule.
A checkbox tells you it finished
Push, email, or Slack. Do not write this into your prompt — it is already built in.
Write a prompt that survives
The part everyone gets wrong the first time.
There is nobody there to answer a follow-up question.
Five rules for an unattended prompt
- Say exactly what to look at — it remembers nothing from your last conversation.
- Say what it must never do — name the destructive verbs out loud.
- Say what to do when unsure — skip, do not guess.
- Define done — in a form you can read in five seconds.
- Tell it to fail loudly — nobody is awake to approve a clever workaround.
The inbox triage prompt
Every run, triage my Gmail inbox and prepare replies I can approve.
Read unread mail from the last 24 hours. Ignore anything already read or archived, and anything in Promotions, Social, or Updates.
Decide which ones need a reply from me personally. Skip newsletters, receipts, automated notifications, and anything a reply would not change.
For those that do, create a Gmail draft. Match my tone: direct, warm, two to five sentences. If it asks something the thread cannot answer, mark a placeholder like [CONFIRM: date] rather than guessing.
Never send, archive, delete, or label anything. Drafts only. If unsure whether something needs a reply, skip it — a missing draft costs me nothing, a wrong one costs me trust.
Finish with one line: "X drafts ready, Y skipped", then one line for each draft: sender — what they want.
Success means every email that needed a reply has a draft waiting, nothing was sent, and I can act on your summary without opening my inbox. If you cannot reach Gmail, stop and say so. Do not work around it.
What it can reach
You give the keys to something that will not ask.
Four things to understand
- It acts as you — drafts, messages and commits carry your name.
- It never asks permission — a run has no approval prompts at all.
- Claude attaches every connector by default — remove what it does not need.
- It can change things, not just read them — connected tools send, edit, delete.
A green run means it finished — not that it worked.
Every run starts from zero
- A routine remembers nothing — no notes from last night.
- Claude copies your project fresh each time — what it wrote last run is gone.
- So keep the memory somewhere real — a tracker, a spreadsheet, a doc.
- Re-check, do not recall — find it fresh, compare against that list.
- Now a skipped run costs you nothing.
Three of them is a system
Each run opens as a full session you can read, continue, or correct.
Ten more jobs for the same machine
No code needed
- Morning inbox triage
- Meeting prep before your first call
- Weekly numbers check from a spreadsheet
- Competitor and market watch
- Turn email promises into tasks
If you have a project
- Nightly review against your own checklist
- Dependency and security sweep
- Find docs that no longer match the app
- Post-deploy smoke check
- Turn an error alert into a draft fix
What we covered
- A routine is a saved prompt plus a trigger. It runs in the cloud without you.
- Name it, write it, pick a time — a code repository is optional.
- Write for a capable stranger who cannot ask you a question.
- Strip the access it does not need before you save it.
- Read the runs. Green only means it finished.
Before we meet again
Pick one thing you do most mornings that a stranger could do from written instructions.
Write it using the five rules — especially "what to do when unsure".
Set it to run daily, turn notifications on, and leave it alone for three days.
Then open all three runs and read what it actually did.
Now go give the thing you do every morning.
It will be ready when you wake up.