Verdict: Claude Routines are Anthropic's April 2026 feature that turns a saved Claude Code prompt into an unattended cloud agent. For a small business, they make sense only for work that repeats daily or weekly, has predictable inputs, and needs judgment — such as curating a morning news brief, drafting inbox replies, or scoring leads. Static data-moving tasks are usually cheaper and safer as database edge functions or cron scripts. Use Routines as the brain, not the plumbing.
Last verified: 2026-06-18 · Best for: recurring judgment work · Not for: one-off, creative-first, or static ETL jobs · Daily caps: Pro 5, Max 15, Team/Enterprise 25
What Claude Routines actually are
A Routine is a saved Claude Code configuration: a self-contained prompt, one or more code repositories, a set of MCP connectors, and a trigger. Anthropic runs it on its own cloud infrastructure, so your laptop can be off and the workflow still executes (Claude, 2026). Triggers can be:
- Scheduled — hourly, daily, weekdays, or weekly.
- API/webhook — a per-routine HTTP endpoint with a bearer token.
- GitHub event — pull request or release activity on a connected repo.
When a trigger fires, Anthropic starts a fresh Claude Code session with your configuration. It reads files, calls connectors, runs commands, and either commits to a branch, opens a PR, posts to Slack, sends an email, or writes to a database — whatever the prompt describes. There is no human-in-the-loop approval step, so the prompt has to be production-ready before you ship it.
This is different from a local scheduled prompt, which only runs while your machine is open. It is also different from a traditional cron job or edge function, which executes the same fixed path every time. A Routine is agentic: it can reroute when a source is down, try an alternative search, or adjust the output format based on intermediate results.
The three-question gate: should you automate this?
Not every workflow should become a Routine. The video creator's framework is a useful filter, and it matches what production operations teams already do. Ask three yes/no questions. Only automate if all three are yes.
- Does it repeat on a predictable cadence? Once a quarter is not a routine. Daily inbox triage, weekly competitor monitoring, or nightly log review is.
- Are the inputs predictable without you? The data source must exist before the routine runs — emails, sign-ups, RSS feeds, API metrics, database rows. If the routine needs your private context to know what to look for, bake that context into a skill or memory file instead.
- Does the failure mode have a fallback? If the routine fails, what happens? A good fallback is "skip that source and continue," "retry twice then ping me on Slack," or "write a failed-run row to the database." A bad fallback is "send a nonsensical email to a customer."
Most workflow ideas fail at least one of these gates. That is fine. The point is to avoid a dashboard full of brittle automations that need more babysitting than the original manual task.
Routine vs. static function: know which layer owns the work
A common mistake is using a Routine for work that should be a database edge function, cron job, or no-code workflow. Use this rule:
| Task type | Use this | Example |
|---|---|---|
| Same path every time, no judgment | Edge function / cron / n8n / Zapier | "Email me yesterday's sign-ups" |
| Needs reasoning, adaptation, fallback | Claude Routine | "Read my inbox, draft replies, flag anything sensitive" |
| Fetch + store raw data | Edge function or ETL script | Sync YouTube stats into Postgres |
| Analyze that data and decide what to do | Routine | Score videos and alert on the top three |
For example, syncing Meta ad spend or YouTube analytics into a database is plumbing. It should run cheaply and deterministically through a Supabase Edge Function or n8n workflow. Turning that sync into a Routine burns Claude usage limits on work that does not need intelligence. The Routine should sit one layer above: it reads the already-synced data, compares it to targets, and writes a recommendation.
The spec-sheet method: write the job description before you build it
Do not open Claude and start prompting. Build a one-page spec sheet first. This is the difference between a hobby automation and a production routine.
Include these six items:
- Goal — one sentence describing the outcome.
- Trigger — when or how it fires.
- Inputs — data sources, skills, memory files, or credentials it needs.
- Steps — the exact sequence the agent should follow.
- Output — where the result lands and in what format.
- Failure mode — what to do when a source, API, or connector fails.
Example spec: "Morning Signal" brief
- Goal: Every weekday at 7:30 a.m., send an email with 3–5 AI/tech stories worth my morning coffee, filtered by my audience and product focus.
- Trigger: Scheduled, weekdays, 7:30 a.m. local time.
- Inputs: Hacker News, TechCrunch, The Verge, a
preferences.mdskill describing my audience, an email template skill. - Steps:
- Fan out to sources and fetch recent posts.
- Normalize titles, URLs, and timestamps.
- Deduplicate against yesterday's picks.
- Score each item against my audience profile.
- Render the top 5 in an HTML email template.
- Send via email connector.
- Log the run and the selected URLs to a database row.
- Output: Email in my inbox; log row in Postgres.
- Failure mode: If one source is unreachable, continue with the others. If the email connector fails, retry once then post a Slack alert. If no items score above threshold, send a "light news day" message rather than a blank email.
This level of detail prevents the routine from making up its own interpretation of "important news."
How to create a Routine in practice
Routines are available to Claude Code users on Pro, Max, Team, and Enterprise plans with Claude Code on the web enabled (Claude, 2026). The two main creation paths are:
- Manual in the web UI — go to
claude.ai/code/routines, choose local or remote, name the routine, paste the prompt, add triggers, and attach connectors. - Via Claude Code CLI — use
/scheduleconversationally, or open a Claude Code session and ask it to build the routine from your spec sheet. It will create the necessary skill files and connector config, then deploy the routine.
When you run a new Routine, treat the first few runs as training runs. Open the Claude Code session it creates, inspect the reasoning trace, and refine the prompt or skills. The goal is not to get a perfect output on day one; it is to make the routine's failures predictable and easy to correct.
Local vs. remote
- Local Routine: Runs on your machine. Can access local files and keeps data off Anthropic's servers. Stops when your laptop sleeps.
- Remote Routine: Runs on Anthropic's cloud. Required for 24/7 schedules, API triggers, and GitHub events. Use this for production business workflows.
For almost any small-business automation, choose remote.
Small-business workflows that fit Routines
Here are workflows that pass the three-question gate and match how a solo founder or small team actually works.
1. Daily inbox triage and draft replies
Why it fits: Repeats daily, inputs are predictable (emails), and the fallback is easy (only draft replies; never send without review). The routine can use a Gmail MCP connector, read your business context from a skill, and produce draft replies inside a shared Notion page or your inbox.
2. Weekly competitor and market signal brief
Why it fits: Repeats weekly, sources are public web pages or RSS feeds, and the output is an internal email. The judgment is in scoring what matters to your positioning, not just scraping headlines.
3. Nightly lead scoring and routing
Why it fits: Inputs are new form submissions or sign-ups in your database; the routine reads them, scores fit, and routes hot leads to a Slack channel or CRM task. Static edge functions do the sync; the Routine does the ranking.
4. GitHub PR review assistant
Why it fits: Triggered on PR open, the routine runs your security and style checklist, leaves inline comments, and flags anything that needs a human. This is judgment work with a safe failure mode: comments can be wrong, but they do not merge code.
5. Content pipeline assistant
Why it fits: Triggered after a new video or blog post is published, the routine drafts social posts, newsletter blurbs, or thumbnail copy based on the transcript. The human approves before anything goes live.
Cost and limits you need to model
Routines share your Claude Code subscription usage, and they have their own daily run caps (Claude, 2026):
| Plan | Monthly price | Daily Routine cap | Best for |
|---|---|---|---|
| Pro | ~$20 | 5 runs/day | Solo founder testing one or two automations |
| Max 5x | $100 | 15 runs/day | Heavy daily use across multiple workflows |
| Max 20x | $200 | 25 runs/day | Power user or small team running many routines |
| Team/Enterprise | $25–$30/seat+ | 25 runs/day | Shared org workflows |
Prices and caps are volatile; confirm at claude.com/pricing before you commit. Beyond the subscription, factor in:
- API costs for connectors. Sending email via Resend is free up to 3,000 emails/month, then $20/month for 50,000. A daily personal brief is effectively free.
- Database and hosting. A Supabase Pro project starts around $25/month and can host the data your Routines read and write.
- No-code automation alternative. n8n cloud starts at €20/month for 2,500 executions. If your workflow is mostly data movement with simple branching, n8n is usually cheaper than Claude Routines. Use Routines for the step that actually needs reasoning.
What this means for you
If you run a small business or solo operation, start with one Routine, not ten. Pick the task that eats a small, fixed slice of your morning every day — inbox triage, a signal brief, or lead routing. Write the spec sheet first. Build the Routine in Claude Code, run it manually a few times, then set the schedule. Keep a human approval gate on anything that touches customers or money.
The real leverage is not that Claude "runs your entire business." It is that you stop spending your best hours on repetitive judgment work that a constrained, well-specified agent can do first-draft for you. You remain the editor; the Routine is the intern.
FAQ
Q: What is a Claude Routine? A: A saved Claude Code configuration — prompt + repos + MCP connectors + trigger — that runs autonomously on Anthropic's cloud. It can fire on a schedule, via an API/webhook call, or on a GitHub event.
Q: Do I need a paid Claude plan to use Routines? A: Yes. Routines are available on Pro, Max, Team, and Enterprise plans. They also require Claude Code on the web to be enabled.
Q: Should I use a Routine or a Zapier/n8n workflow? A: Use n8n, Zapier, or a database cron for deterministic data movement. Use a Claude Routine when the step needs judgment, adaptation, or fallback reasoning that you cannot express as fixed branches.
Q: Can a Claude Routine replace human decision-making? A: No. A Routine should handle first drafts, scoring, flagging, and routing. Any action that affects customers, money, or compliance should keep a human approval step.
Q: What makes a Routine fail in production? A: The most common failures are vague prompts, missing failure modes, connectors with stale OAuth tokens, and trying to automate work that is not actually repetitive. A good spec sheet prevents most of these.
Q: How many Routines can I run per day? A: As of June 2026, Pro users get 5 Routine runs per day, Max 5x gets 15, Max 20x and Team/Enterprise get 25. These limits are in addition to your normal Claude Code usage pool and can change during the research preview.
Discussion
0 comments