What Claude Code Can Replace: A Worked Example

Claims about AI saving weeks of work are everywhere, and most of them come with no detail. This post is a plain look at what Claude Code can take on in a small business, what it does well, and where you still have to step in.

It contains no before and after stopwatch numbers, because none were measured for this post. Any example here is an illustration.

The gain is not typing less. It is that jobs that needed a specialist or a long block of time can be started by describing them, then reviewed and corrected in rounds.

What it can take on

Anthropic describes Claude Code as a tool that reads files, edits them and runs commands. The Halftime lesson shows jobs such as these:

• Dashboards built from real numbers.
• Test setups for software projects.
• Landing pages and website changes.
• Planning and drafting social posts.
• Sorting messy files and folders.
• Drafting replies and slides.

The Halftime lesson on Claude Code covers this kind of work, such as cleaning up a messy computer, building a page and drafting inbox replies. It is a starting point, not a promise of any result.

Why it works for this

Three habits turn the tool from a toy into a working part of the week.

It sees the real files. It does not guess what your project looks like. It reads it. See the tutorial for a first session.

It can carry notes between sessions. Each session starts fresh, but a CLAUDE.md memory file holds the business, the voice and the decisions, so each session starts informed. See giving AI a memory.

It can keep rules. Corrections you save become written rules. They are context rather than hard locks, so check that they are followed. See the brief once method.

What it does not do for you

Be honest about the limits.

• It does not know what is worth building. You choose the goal.
• It can be wrong with confidence, so someone has to check the output.
• It can touch things you did not intend if you give it loose instructions.
• It does not replace judgement about customers, money or risk.

Time saved is also not free. You spend time briefing, reviewing and correcting. The gain comes when that cost is smaller than doing the job from scratch, which is most often true for repeated, well defined work.

How to test it on your own work

Do not take anyone's word, including this post. Run a small experiment.

1. Pick a real jobSomething you would normally spend a few hours on, with a clear finish line.
2. Time your usual wayNote roughly how long it usually takes you, or took last time.
3. Brief it properlyUse the five part brief.
4. Count everythingInclude briefing, waiting, reviewing and fixing, not only the part where it works.
5. Compare honestlyIf the total is lower and the quality is acceptable, do the next job. If not, learn why.

Where the biggest wins tend to be

Jobs that are repeated, easy to check and easy to undo. A dashboard you can compare against your numbers. A page you can open and look at. A folder you can inspect. Pick those first. Our guide to what to automate first shows how to score them.

The caution

If a post, including a lesson, promises a fixed amount of time saved for everyone, be careful. Results depend on your skills, your tasks and how well you brief. Treat the numbers as examples to test, not guarantees.

A worked illustration of the method

Say you run a small online shop and want a simple dashboard of weekly orders. The rough shape of the job with an agent might look like this. You export your orders to a spreadsheet. You ask Claude Code to read it, describe what columns it found, and propose a one page dashboard. You say which numbers matter most. It builds a first version as a page you can open. You compare its totals against the spreadsheet by hand for two or three weeks. You ask for corrections. Then you save the rules so the next dashboard follows the same layout.

This is an example, not a record of anyone's results. The point it shows is where your time goes: describing, checking and correcting, instead of building from nothing.

Questions to ask about any AI success story

Who did the work, and how skilled are they? What exactly was measured, and over what period? Was checking and fixing included in the time? Could someone else repeat it? A good story answers these. A weak one just gives a big number.

What to measure on your own jobs

Track three things for a month. The time you spend on the job end to end, how many corrections the first draft needed, and how many mistakes reached a customer. If the first falls, the second falls over time, and the third stays at zero, you have something real. If not, the tool may fit a different job.

Questions people ask

Does Claude Code really save weeks of work?

No one can promise that. Results depend on the job and your briefing. Test it on one real task and count all your time.

What should I use it for first?

Repeated, well defined work that you can easily check and undo.

Do I still need to check what it makes?

Yes. It can be wrong while sounding sure. Review everything that matters.

Is this a replacement for hiring people?

It can take on parts of jobs, but judgement, customer care and responsibility stay with you.

Next step

See how it is set up and used in Lesson 1.1, and decide with your own test.