October 8, 2026Authority AI Editorial

Turn a Weekly Job Into an Automation Your Agent Runs While You're Away

A row of glowing glass stations joined by a violet light path, with a small gate between each one, ends in a neat finished document resting on a dark desk at dawn.

Some jobs come back every week: the Monday competitor check, the Friday client summary, the sweep through a shared inbox. You know exactly how each one goes, because you've done it fifty times.

That familiarity is what makes these jobs good candidates for an automation: a set of steps your agent works through on a schedule, without you sitting there. The aim is simple. You review the result instead of doing every step yourself.

Skills and automations are different things

It helps to keep two ideas apart.

A skill is a written method for one kind of job: what a good result looks like, the steps, the checks and the usual traps. Your agent follows it when you ask for that job. A skill doesn't start on its own. Someone, or something, has to hand the agent the request.

An automation is that something. It decides when the job runs, which skills are used and in what order, and what happens between them. Skills hold the know-how. The automation holds the timing and the order.

So a weekly automation is usually a short chain of skills you already trust, with a start time and a few checks added.

Which weekly jobs to hand over first

Not every recurring job is a good first automation. The best ones share three traits.

The steps are the same each time. If a numbered list of them would still be right next month, it's a candidate.

The inputs are easy to find. A set of web pages, a folder of documents, a shared inbox. If the job starts with "first, remember what Sam said on the phone", it isn't ready.

A mistake is cheap to catch. A draft you read before it goes anywhere is a safe first automation. Anything that sends money, emails customers or deletes things is not.

Anthropic's engineering team, writing about how to build agent systems, recommends "finding the simplest solution possible, and only increasing complexity when needed." That's good advice for your first automation too. Start with one job, a short chain, and a result only you will see.

One skill per step

Take a weekly competitor check. What follows is an illustration of how the job might be split, not a logged run.

Gather. Visit three competitors' pricing pages and their latest posts, and note what's on each.
Compare. Set this week's notes beside last week's and list what changed.
Write. Turn the changes into a one-page summary, most important first.
Check. Match every price and date in the summary against the page it came from.

Each step is its own skill, with its own idea of what good looks like. The gathering skill knows to record the address and date of every page it visits. The writing skill knows the summary is one page and starts with what changed, not with what stayed the same. The checking skill knows that any figure it can't match gets flagged, not smoothed over.

Splitting the job this way keeps each skill short enough to be clear, and shows you exactly which step went wrong when one does.

Anthropic calls this pattern prompt chaining: a task broken into a fixed sequence of steps, each one working from what the step before it produced. Their guidance says it suits work that "can be easily and cleanly decomposed into fixed subtasks." A weekly job you've done fifty times is about as fixed as work gets.

The check between steps

The part that makes an automation safe to leave alone is what happens between the steps.

Anthropic describes adding checks on the steps in the middle of the chain, which their diagram labels a gate, to keep the process on track. In plain terms, a gate is a question the automation asks before it moves on, with a clear answer of yes or no.

For the competitor check, the gates might be:

After gathering: did all three pages load, and does each one show a price? If not, stop.
After comparing: is the list of changes shorter than the whole page? If every line changed, something probably went wrong with the gathering, so stop.
After writing: does every figure in the summary appear in the notes? If not, stop.

"Stop" matters more than it looks. When a step fails, an agent left to itself will often fill the gap with something plausible. A gate turns that quiet guess into a visible halt, with a note saying which step stopped and why. A short message saying a page didn't load beats a confident summary built on a missing price.

Hospitals have spent years on the moment one shift hands a patient to the next, because that's where information goes missing. In a 2014 study across nine hospitals, published in the New England Journal of Medicine, a handoff programme with a fixed structure for what gets passed on was followed by a 23 percent fall in the medical-error rate and a 30 percent fall in preventable adverse events.

In an automation, the gate is that handover: the point where you decide, in advance, what has to be true before the work moves on.

Where the run happens

Claude Code's documentation describes three ways to run a prompt on a schedule, and the difference is mostly about where the work happens.

A cloud task runs on Anthropic's machines, so your computer can be off. It works from a fresh copy of your project rather than your local files, and the shortest gap between runs is an hour.

A desktop task runs on your own computer. It can reach your local files and tools, but your machine has to be on when the time comes.

A loop runs inside a session you have open. It stops when the session closes, and recurring loops expire after seven days, so it suits checks during the day rather than a weekly job.

For a weekly job, choose between the first two by where the inputs live. If the job needs files on your laptop, it runs on your laptop.

Then settle what the run is allowed to do. The documentation on unattended runs is direct: in a scheduled job nobody is available to answer a request for permission, so anything that would normally ask you is refused and the run carries on without it. Decide up front which actions the automation can take on its own. For a first automation, reading pages and writing a draft is plenty.

What to keep a person on

Some parts of a weekly job should stay with you, at least at first.

Anything that leaves the building. The automation can draft the client summary. You send it.

Anything that can't be undone. Archiving emails, changing a shared document, deleting old files. Let the automation propose these as a list, and do them yourself.

Judgement calls. The automation can report a competitor's price drop. Whether you respond is your decision.

Anthropic's guidance on agents makes a similar point: they recommend testing agents thoroughly in a safe, separate setup first, with guardrails in place, and note that agents can "pause for human feedback at checkpoints." The checkpoint here is you, reading the result before anything happens because of it.

Reading Monday's result

A good automation hands back two things: the finished piece of work and a short note on how the run went.

Read the note first: did every step finish, did a gate stop the run, were any figures flagged? If it's clean, read the summary. If not, it tells you which step to look at.

Over the first few weeks, keep a list of what you change by hand. Each item is a fix for one of the skills: a page the gathering step keeps missing, a summary that keeps leading with the wrong thing. The automation itself rarely needs to change. The skills get sharper.

Where to start

Pick one job you do every week that only you read. Write it down as four or five numbered steps. Under each step, write one sentence on what has to be true before the next step can start. Those sentences are your gates.

Run the chain by hand once with your agent and notice where you had to step in. Fix those skills, and only then put the job on a schedule.

If you'd rather start from methods someone else has written down, the AI Ultimate Skill Collection is 3,029 ready-made skills sorted into 12 packs, one for each kind of work, such as coding, research or finance, and it works with Claude Code and other agents that use skills. The Agent Automation Collection is a set of automations written in-house rather than collected, built to chain skills together for the jobs that come back every week, with scheduling, unattended runs and guardrails included. Same folder, same agent, nothing to install.