What would you pay $20 to get automated? That's your use case
What can I automate with Claude, Claude Code or Codex? Use a simple $20 test for picking the right job, six recurring jobs that pass it, what not to automate, the honest cost of running them yourself, and a prompt to start one in Kavela.
What can I automate with Claude? Any recurring job where the inputs are reachable, the output is checkable, and a mistake can be caught before it reaches anyone. The fastest way to pick one: list your weekly chores and ask which one you would pay $20 a month to never do again. That job is your first use case.
How do you decide what to automate with Claude?
Most people start from the tool ("Claude can read my Gmail, what should it do?") and end up automating something that did not matter. Start from your week instead, and run every chore through the $20 test:
- List what you did more than once last month. Not ideas, actual repeats: the Monday report, the inbox sweep, the release notes, the LinkedIn post you keep meaning to write.
- Price each one. Would you pay $20 a month for someone to do it reliably, forever? If the answer is an instant yes, it is worth automating. If you hesitate, it is either too rare or too small.
- Check the inputs. Can an agent reach the material? Your calendar, a repository, a Notion page, public web pages. If the input lives in your head, it fails.
- Check the output. Can you tell in under a minute whether the result is right? A draft post, a changelog, a list of changed prices: yes. "Improve our strategy": no.
- Find the judgement point. Where would you want to say yes or no before it goes out? That is your approval step.
Teresa Torres, who writes about running her work through Claude Code, sorts candidates by a similar test in How to Choose Which Tasks to Automate with AI. She uses AI to "do more of what I'm already good at," "eliminate friction in the tasks that I do often," and "remove what drains me." The payoff she reports is concrete: about 35,000 words a month, "up from about 8,000."
What makes a job a good fit for an AI agent?
A job passes when all five of these hold:
| Test | Passes | Fails |
|---|---|---|
| It recurs | Weekly release notes | A one-time migration |
| Inputs are reachable | Merged pull requests | "What the client meant on the call" |
| Output is checkable | A list of changed competitor prices | A brand strategy |
| The judgement point is clear | Approve each post before it publishes | "Use your judgement" |
| Mistakes are gated or reversible | Gmail drafts, labels | Deleting files, sending invoices |
The last row matters most. An agent that drafts is low risk because you see the draft. An agent that sends, deletes or pays needs an approval step in front of that action, every time.
What can you automate with Claude right now?
Six jobs that pass the test for a small software company, agency or founder. Each links to a page with the steps, the do-it-yourself setup and a prompt you can paste:
| Job | Runs | You approve |
|---|---|---|
| LinkedIn posts from your real week | Weekly | Each post's exact text |
| Competitor analysis: pricing, launches, messaging | Weekly | Nothing to approve: a report to read |
| Email triage for Gmail | Twice a day | Each reply before it sends |
| Release notes from GitHub pull requests | Weekly | The draft before it publishes |
| SEO audit of your key pages | Weekly | Fixes before a pull request opens |
| A daily briefing across your tools | Weekday mornings | Any action it suggests |
Anthropic's own routines documentation lists engineering versions of the same idea: nightly backlog grooming, alert triage, docs drift checks and deploy verification. The pattern is identical: a clear trigger, unattended work, and "a clear outcome."
What should you not automate with Claude?
- Anything you have done once. Automate the second time, when you know what good looks like.
- Decisions with no checkable output. Hiring, pricing, positioning. Use Claude to prepare the material, and keep the call.
- Irreversible actions without a gate. Payments, deletions, messages to customers. An approval step or nothing.
- Engagement on social platforms. Auto-likes and auto-comments get accounts flagged. Posting your own approved content is fine.
- Jobs whose input is your memory. If you cannot write the input down, the agent cannot read it.
What does it cost to run these automations yourself?
Less than it used to, and more than the demos show. Claude Code now has routines that run in Anthropic's cloud on a schedule, with a minimum interval of one hour. Codex has scheduled tasks, and tasks that work in a local project need "the computer on and the app running."
The work that is left is coordination, and it lands on you:
- The approval gap. Routines run "without stopping for approval," so any job that publishes or sends has to be split into a draft run and a publish run you wire yourself.
- Memory between runs. A routine clones your repository fresh each time. Last week's report has to be committed back or stored somewhere else.
- Network and credentials. Routines allow a default list of domains; anything else fails with a 403 until you edit the environment. Tokens expire on their own schedules.
- Session juggling. Six routines across two tools means six places to check, and remembering which session knew what.
That is the productivity paradox: the agent does each job in minutes, and you spend your evenings keeping six of them running. None of it is hard. It is just yours to carry.
How do you start your first automation in Kavela?
Kavela is a project workspace where Kav carries the work between your decisions. You describe the job in one chat, Kav turns it into a task, briefs agents for it, asks for the connections it needs, and brings you only approvals and finished results. Routines run in a cloud workspace, so your laptop can stay closed. Your own Claude or Codex subscription runs on every plan, including Free.
If you already know your $20 job, open its use case above and paste the prompt. If you do not, paste this one into Kav and let it find the job with you:
Help me find the one recurring job I should automate first.
Ask me up to 5 short questions about what I did more than once last month and which tools I use (calendar, email, GitHub, Notion, Slack, LinkedIn).
For each repeat job I name, apply this test: would I pay $20 a month to never do it again, can you reach the inputs, can I check the output in under a minute, and where should I approve before anything goes out.
Recommend one job, explain why in three sentences, and show me the routine you would set: schedule in <your time zone>, inputs, connections needed, and the approval step.
When I say yes, ask me to connect what it needs, run it once so I can check the first result, then set the routine.
You still sign in to each service once and approve the first result, so it is not zero setup. For most of these jobs it is about 10 minutes when you already have the accounts.
Frequently asked questions
What is the easiest thing to automate with Claude first?
A weekly report you already write by hand, such as release notes or a competitor check. The inputs are easy to reach, and you can tell at a glance whether the output is right.
Can Claude automate tasks without me writing code?
Yes. Claude Code and Codex both take plain-language instructions, and Claude Cowork is built for non-coding work. Writing the instructions well, with inputs, output and an approval step, matters more than code.
Is Claude or ChatGPT better for automation?
For recurring jobs, the scheduling and the connections matter more than the model. Claude Code routines and Codex scheduled tasks both run jobs on a schedule; compare where they run, what they can connect to and whether they can pause for your approval.
How much does it cost to automate a task with Claude?
Routines draw down your Claude subscription the same way interactive sessions do, so the cost is mostly your plan's usage. The larger cost is your time setting up and maintaining each job, which is what the $20 test helps you justify.
For a deeper look at handing over whole outcomes instead of single prompts, read how to use AI more efficiently, and for scheduling, Claude calendar.
Stuck on setup? Ask in the Kavela Discord.
