Claude Code Routines: How to Run Your Work Without Being There
You open the laptop, Claude works. You close the laptop, Claude stops. Every task waits for you to start it, watch it, approve it, save it. The bottleneck on your AI-powered business isn't the model. It's you being there for it to happen. Routines are saved Claude Code sessions that run in Anthropic's cloud on a schedule, a webhook, or a GitHub event... autonomously, while you sleep.
Marlon Brand
Founder, Undeniable · Last updated April 2026
The Supervision Tax
You use Claude Code for real work now. You've seen it draft client reports, clean up copy, summarize calls, refactor a landing page at midnight. You trust it enough to let it run longer and longer sessions.
Every session still starts with you. Every task still needs you present. Open the terminal, paste the context, describe the job, watch it run, review the output, save it somewhere. Close the terminal. Tomorrow you do it again.
Count the hours. A consultant or small agency owner running two to four Claude Code sessions a day at 30 to 60 minutes each is spending 2 to 3 hours of direct attention on AI work. Not thinking. Not strategy. Supervision... watching a tool do a thing.
At your rate... $100, $150, $250 an hour... the math lands somewhere between $50,000 and $200,000 a year of your presence as the bottleneck. Pick your own number. The pattern holds at every rate. Your business has an AI infrastructure that only functions when you're there to run it.
That's the Supervision Tax. The third one.
The Reset Tax: context re-explained every session. The slash commands guide covers it. The Format Tax: video, audio, and images translated before AI can touch them. The Gemini guide covers that one. The Supervision Tax: every task requiring you to start and watch it.
Three taxes. One shape. Each one is the hidden cost of treating AI like a tool you use instead of infrastructure you build.
Slash commands killed the first tax. Multimodal killed the second. Routines kill the third.
What Routines Actually Are
A routine is a saved Claude Code session that runs without you. You write the prompt once, attach a trigger, walk away. The session runs in Anthropic's cloud on a schedule, or when an API call hits a URL, or when something happens in a GitHub repo. You wake up to completed work.
This is structurally different from anything else in Claude Code. When you open the CLI and type a prompt, Claude runs on your machine, in your terminal, with you watching. When you run /loop or a desktop scheduled task, the work still happens locally... your laptop has to be open, the app has to be running, your attention isn't required but your hardware is.
A routine is none of that. It runs on Anthropic-managed cloud infrastructure. Your laptop is closed. You're asleep. The session still executes, reads repos, calls connectors, opens pull requests, posts to Slack. You see the output when you wake up.
This is Claude Code as infrastructure, not Claude Code as a tool.
One caveat. Routines are in research preview as of this writing. Behavior, limits, and the API surface may change. What doesn't change is the category. Autonomous AI that runs without human supervision is the direction the entire Claude Code product is moving. The specific features will evolve. The shift in how you work is permanent.
New to Claude Code? This guide is easier if you've spent time in Claude Code already. If you haven't, start at Learn Claude Code... you can still read this guide for the shape of what's possible.
The Three Triggers
A routine needs something to start it. Anthropic gives you three options. They're not interchangeable... each one fits a different kind of work.
Schedule Triggers
Run a routine on a recurring clock. Hourly, daily, weekdays, weekly. Custom cron if you want it. This is the trigger for work that happens on a cadence regardless of what else is going on... Monday morning client reports, nightly content audits, weekly drift checks.
API Triggers
Give a routine a dedicated URL and a bearer token. Anything that can make an HTTP POST can trigger the routine. A form submission on your site, a monitoring alert from your infrastructure, a webhook from Stripe when a subscription upgrades. You pass in context as a text field, the routine picks it up. This is the trigger for work that reacts to specific events.
GitHub Triggers
Run a routine when something happens in a repo you've connected. Pull request opened, release published, label added. Filters let you narrow it to specific branches, authors, or labels. This is the trigger for work that belongs in your codebase... copy PR reviews, documentation drift checks, automated style enforcement.
One routine can combine triggers. A copy review routine can run on every GitHub PR, trigger from your CD pipeline after each deploy, and also run nightly to catch anything that slipped through. One prompt, three ways to fire it.
Before routines, every trigger for AI work was you. Now you choose.
Four Routines That Change How You Work
Four routines. Each one kills a specific weekly task that used to need you in the room.
Morning Brief
- Trigger
- Schedule, 7:00am weekdays
- Connectors
- Slack, Google Calendar
- Output
- Pulls unread priority threads, reads today's calendar, drafts a 5-bullet summary, posts it to a private channel.
You open Slack at 7:30am (or your email inbox, if Slack isn't your hub) expecting chaos. Instead you see a pinned message from yourself, generated 30 minutes ago. Three unread threads need a response. Two meetings today, here's who's attending, here's what they sent last week. One proposal still waiting on your review since Tuesday.
Before this routine, your first 20 to 45 minutes of the day is reconstruction. What happened overnight. What's urgent. What's noise. After this routine, you start the day already caught up.
At 45 minutes saved per day, five days a week, 50 weeks a year, that's 187 hours. At $150 an hour, that's $28,000 you get back. Use your own rate if it's different. One routine.
Copy PR Review
- Trigger
- GitHub event, pull_request.opened, on content repos. (If copy doesn't flow through GitHub, fire the same pattern via an API trigger from Google Docs comments or your CMS.)
- Connectors
- None (the repo and your voice spec are enough)
- Output
- Reads the diff, checks against COPYWRITING.md, posts inline comments for violations.
When anyone opens a pull request in a content repo, Claude reads the diff, checks it against your COPYWRITING.md voice spec, and posts inline comments for any violations... catching em dashes, flagging banned phrases, and highlighting off-brand sentence patterns before a human ever opens the PR.
This is the routine that makes voice consistency scale beyond one writer. Before, every new piece of content needed your best writer (or you) to read it and catch the AI-generated patterns. After, the routine catches the mechanical violations before a human ever opens the PR. Human review focuses on judgment: is this the right angle, does this argument hold. Not spell-check.
The prompt references your actual COPYWRITING.md file, loaded from the repo on every run. When you update the voice spec, the routine updates with it.
Client Report Draft
- Trigger
- Schedule, Monday 6:00am
- Connectors
- Meta Ads, GA, CRM, or your coaching notes folder
- Output
- Pulls last week's data or session notes, compares to the previous week, drafts a rough weekly artifact in your template.
Works for any client you owe a weekly artifact to... a Meta ads performance report, a 1:1 coaching progress snapshot, a fractional CMO update. The shape is the same. Four clients, 45 minutes each, three hours gone before you've done any real thinking.
The routine doesn't replace your judgment... the draft is still rough, the insights are still yours to add. What it replaces is the data-gathering and template-filling that doesn't require judgment.
Twenty minutes of polish per report instead of 45 of assembly. Four clients a week, 50 weeks a year, that's 83 hours recovered. Your Monday mornings become strategic instead of clerical.
Lead Webhook Handler
- Trigger
- API endpoint, POST from your form tool
- Connectors
- Email, CRM
- Output
- Reads applicant answers, drafts a personalized reply as a Gmail draft, logs the lead in CRM.
When a high-intent lead submits your application form, the form tool fires a webhook to your routine's URL. The routine reads their answers, drafts a personalized response tailored to what they wrote, saves it as a draft in your email, and logs the lead in your CRM with enrichment notes.
You reply to new applications in 12 to 48 hours today. A routine replies at minute two, while you're still asleep. The draft is ready in your inbox when you wake up.
The reply isn't sent automatically. It's a draft. You open your email, read what the applicant wrote, read what Claude drafted, adjust two sentences, hit send. Ten seconds instead of ten minutes. Your first-touch reply time drops from 12 to 48 hours to under one hour.
Four routines. One prompt each. Every week, they return hours you used to spend being the manual trigger.
The Prompt Is The Product
Interactive Claude sessions forgive a vague prompt. You see the output and correct it mid-stream. Ask again. Clarify. Iterate. The conversation does the heavy lifting.
Routines don't have that. A routine runs autonomously. No permission prompts, no clarifying questions, no human in the loop until the work is done. Whatever the prompt tells Claude to do is what Claude does. If the prompt is ambiguous, the output is ambiguous. If the prompt doesn't define success, Claude guesses at success.
The prompt itself becomes infrastructure. It has to be self-contained, explicit about what to do, and explicit about what “done” looks like.
Vague prompt (works interactively, fails in a routine)
Self-contained prompt (works in a routine)
The second version reads like a work order. Scope defined, rules referenced, outputs specified, edge case handled. It works because it leaves nothing for Claude to interpret.
Writing good routine prompts is a skill. It's also the highest-return version of prompt engineering you'll ever do... because every run of every routine compounds on the quality of that one prompt. A sloppy prompt in an interactive session costs you five minutes. A sloppy prompt in a daily routine costs you every day, forever, until you fix it.
Spend an hour writing the prompt. Spend another hour testing it by running the routine manually a few times. Then let it run.
The Security Ledger
Routines act as you. Every commit a routine pushes is attributed to your GitHub account. Every Slack message appears under your name. Every Linear ticket it creates is assigned to you in the audit log. This is powerful and it's a liability. Handle it with the same care you'd give any automation that speaks for you.
Three safeguards worth implementing before your first routine goes live.
Scope repositories tightly
A routine only needs access to the repos it actually works in. If the routine reviews copy PRs in your content repo, don't give it your infrastructure repo. Each routine, each repo, one purpose. Don't create an “everything” routine with access to every repo you own.
Keep unrestricted branch pushes disabled
By default, Claude can only push to branches prefixed with claude/. This is the right default. Leave it on. It means a routine can open pull requests, but it can't accidentally push directly to main or a protected production branch. When you review what the routine did, you review it as a PR. Human approval before merge.
Scope connectors to what the routine needs
When you create a routine, every connector you've authorized is attached by default. Strip that down to the specific connectors the routine uses. A copy review routine doesn't need access to your Stripe connector. A morning brief routine doesn't need access to your GitHub repos. Remove what isn't needed on every routine.
One more point that's easy to forget. The prompt itself is a security boundary. If your prompt says “post a summary to Slack,” Claude can post to Slack. If your prompt says “also send an email to the client,” Claude can send email. Be explicit about what the routine is allowed to do and what it's explicitly not allowed to do. Treat the prompt like permissions policy, not just instructions.
When Routines Beat Interactive Claude
Not every AI task should be a routine. Some work still belongs in an interactive session. The decision frame is short.
Make it a routine if ALL are true
- The task is repeatable. It has a shape that holds across instances.
- Success is clear. You can describe what “done” looks like in one sentence.
- No mid-run judgment required. The work doesn't need you to make a call halfway through.
Keep it interactive if ANY are true
- The task involves strategy or new thinking. Brainstorming, positioning, creative direction.
- Success depends on context only you have. The right output changes based on what a client said yesterday.
- The output is sensitive enough that a human must see every version before it ships.
The wrong way to think about this is “routines are better because they're automated.” Routines and interactive sessions do different jobs. Judgment work stays interactive. Repetition becomes a routine.
Most AI workflows have both kinds of work tangled together. The skill is separating them. Pull out the repetitive, rule-bound parts. Make those routines. Keep the judgment-heavy parts interactive. Your setup ends up with a handful of routines running on their own, feeding clean work into the sessions where you actually think.
Routines run. You think.
Start With One
Don't build five routines this week. Build one. Watch it run for five days. Fix what breaks. Then add a second.
Start with the morning brief. It's the lowest-risk routine to build first... it only reads, it only posts to a private channel, it can't do any damage if the prompt is wrong. And it saves time from day one.
The 15-minute setup
- 1Open claude.ai/code/routines. Click New routine.
- 2Name it “Morning Brief.” Write the prompt. Be explicit: which Slack channels to check, what counts as priority, how many items to summarize, where to post the output.
- 3Select any repo (it just needs one to clone). Leave unrestricted branch pushes off.
- 4Attach the Slack and Google Calendar connectors. Remove everything else.
- 5Add a schedule trigger. Weekdays at 7:00am.
- 6Save. Click Run now to test it.
If the first run's output is close to what you wanted, let it run for a week. If it's off, edit the prompt and re-run. Iterate until the output is useful... then walk away and let it run.
The 30-day ladder to four routines
By the end of month one, you have four routines running. They don't add a separate bill. Runs count against your existing subscription usage. They save five to ten hours a week of direct attention... not every hour of the Supervision Tax, but most of the repetitive ones. Call it half the tax, paid off. And you stopped being the bottleneck on your own AI infrastructure.
That's the Supervision Tax, cut in half.
Build the first routine this week. Let the second find you.
If you want the shape of a full self-running infrastructure walked through live, that's what the cohort does. And if the ladder from one routine to done-for-you infrastructure feels like more than a weekend project, that's the bridge to what we build for clients at $500+/day spend.
Common Questions, Straight Answers
// Free download
The Claude Code Routines Cheatsheet
The four starter routines, the trigger types, the prompt pattern, and the security checklist on a single page. Print it, pin it above your monitor, or keep it open in a tab. No email required.
Download the Cheatsheet (PDF)↓More from the Lab
How We Cut Cost Per Lead from $16 to $5.43
The creative strategy framework behind it. Same budget, same targeting. Different creative.
Read →--meta-adsMeta Ads Credit Card Deadline: What to Do Before April 1st
How to switch billing, what it means for cash flow, and the exact message to send each type of client.
Read →--lead-genFacebook Lead Forms vs Landing Pages
Which funnel setup actually converts and when to use each.
Read →--ai-toolsClaude vs ChatGPT for Coaches and Consultants
Which AI actually does the work? Architecture, persistence, and compounding compared.
Read →--ai-toolsControl Your Computer From Your Phone Using Claude
Step-by-step Remote Control setup. No coding required. 10 minutes.
Read →--claude-codeClaude Code Slash Commands That Actually Matter
The slash commands that build persistent context instead of disposable chats. /init, /memory, /compact, /plan, and custom skills.
Read →--geminiGemini Multimodal: The AI That Sees Your Work
When to use Gemini instead of Claude or ChatGPT. Video, photos, recordings, and image generation for coaches and consultants.
Read →--email-deliverabilityGet Out of Spam — Or Keep Losing Revenue in Silence
Why your emails land in spam and the infrastructure framework for fixing it.
Read →--ai-motion-controlAI Motion Control: Cinematic Video with Higgsfield
Create cinematic AI video using Higgsfield. No camera crew, no budget.
Read →--free-toolClean Copy — Make AI Text Sound Human
Free browser tool that strips AI tells from your copy. Paste, pick a mode, get human output.
Try it →--free-courseLearn Claude Code — Free 5-Lesson Video Course
Set up VS Code with Claude Code, build a landing page, and deploy live with Vercel. Five lessons, no coding experience required.
Watch free →--ai-toolsSix Steps to a Claude That Actually Works for Your Business
Set up Claude with memory, projects, and custom writing styles so it works like a trained team member. Six setup steps that compound.
Read →--ai-toolsHow to Switch from ChatGPT to Claude (And Why It Matters for Your Business)
Where Claude outperforms ChatGPT for business work, and how to bring your ChatGPT context into Claude in about twenty minutes.
Read →--ai-toolsClaude Cowork: Stop Asking. Start Delegating.
Getting started with Claude Cowork. Covers Global Instructions, autonomous tasks, and saved Skills. For business owners who want to delegate, not just chat.
Read →Stop Running Your AI. Start Operating It.
We build the full self-running infrastructure underneath all this for one client per quarter. The routines, the custom skills, the CLAUDE.md files, the whole system.
// One client per quarter
Apply to Work With UndeniableNo commitment. No pitch deck. Just a straight conversation about what's possible.