Claude Code is Anthropic’s tool for getting work done inside a folder of files. It began in early 2025 as a command-line tool for programmers, and it still looks like one. It now also runs as a desktop app, in the browser and inside code editors, and the thing it is good at has quietly outgrown programming: reading a pile of documents, following written procedures, using the tools a business already pays for, and writing back what it did.
That makes it interesting to a business owner for a reason that has nothing to do with software. Most operations already run on files and written instructions: the onboarding checklist, the proposal template, the brand document, the spreadsheet of leads, the standard reply to a common question. Claude Code works on exactly that material, the way a capable new hire would with the manual open.
This piece is about where that is genuinely useful, where it is the wrong tool, and how we decide between the two.
What Claude Code actually is
Strip away the terminal and it is four things.
A folder it can work in. You point it at a directory. It reads the files there, writes new ones, edits existing ones and runs commands. For a business that folder holds procedures, templates, records and past work rather than source code.
Written instructions. A plain text file at the top of the folder tells it how this business works: who the customers are, what the voice sounds like, what it must never do, where things live. Repeatable jobs get their own instruction files, which Anthropic calls skills. A skill is a procedure written in plain language: “to draft the weekly client update, read these three files, follow this structure, flag anything that looks like a number you cannot source.”
Connections to the tools you already use. Through a standard called MCP it can read and write Notion, Airtable, Google Sheets, Slack, Gmail, calendars and most things with an API. It does not replace those tools. It works inside them.
Memory and schedules. It can keep notes between sessions about what changed and what you corrected, and routine jobs can run on a timer without anyone opening a window.
The result is not a chatbot. It is closer to an assistant with a desk in your file system.
What it is not
It is not the right tool for everything with the word AI near it, and the failures we see come from pretending otherwise.
It is not a workflow engine. If a job is “every two hours, check for replies and mark the record”, that is fixed steps with no judgment in them. A tool like n8n or Make runs that cheaper, faster and more predictably, and it will still be running at 3am in six months. Claude Code can do it, in the sense that a surgeon can take out the bins.
It is not a database or a system of record. It reads and writes to those. The record of truth stays in the tool your team already trusts.
It is not a website chatbot or a phone agent. Customer-facing conversation in real time needs a different kind of system, and voice needs a voice platform.
And it is not something to leave alone with your money, your contracts or your customers. Anything that leaves the business or changes a record other people rely on gets a review step until the pattern has earned trust.
Where it earned its place in real work
The pattern in the systems where it fit is the same every time: messy inputs, a lot of shared context, judgment within limits, and a person deciding at the end.
Our own operation. Aevolve runs on it. The folder holds the positioning, the voice rules, the lead database and a set of skills: drafting LinkedIn posts from a research note, drafting comments, preparing outreach messages, keeping the pipeline in Notion current, running a research brief. Every output is a draft that a person reads before anything leaves. The instructions have been edited hundreds of times, in plain English, as the business learned what works.
A content system for a media company founder. For Avision Digital we built a system that collects videos that already performed in his niche, scores new ideas against his audience and his offer, and drafts scripts from his own brand documents. Airtable holds the ideas and the scores. He decides what gets filmed. The first batch of scripts came back generic. The fix was not a new tool, it was a better instruction file and a tighter set of source documents, which is the kind of change the owner can make.
Weekly LinkedIn drafts for a founder. Drafts of posts and images land in Airtable for the founder to edit. What they change feeds back into the instructions, so the drafts drift toward their voice over time.
None of these jobs could have been a fixed workflow. All of them produce drafts, not final actions.
Where we did not use it
The honest list matters more than the success list.
For Gizoom’s outreach engine there is no AI at all. The job was finding companies, verifying every email address, keeping one clean list so nobody is contacted twice, sending from warmed-up mailboxes and following up until someone replies. Every step has a rule. Make and Smartlead run it, and a person takes over the moment a warm reply arrives.
For Avision’s booking automation exactly one step needs AI: reading a booking email that is written differently every time and pulling out the client and the property. That single step sits inside a Make scenario. Everything after it, folders, share link, job, team message, is plain automation.
For a hotel spa’s phone receptionist, the system is a voice platform wired to n8n and the spa’s booking software. Claude Code had no role in it.
Choosing the simpler thing is most of the job.
How to tell if a job belongs to Claude Code
We mark every step of a workflow with one of six answers: connect the tools, automate it, use AI, leave it human, build focused software, or simplify the process. A step earns the AI answer when its input is messy, when it needs judgment inside clear limits, or when the right action depends on context a rule cannot hold.
Within the AI answer, Claude Code specifically is the right tool when most of these are true:
- The work lives in documents and text, and there is a lot of it.
- Several related jobs share the same context: one brand, one client list, one set of procedures.
- A person reviews the output before it goes anywhere.
- The procedure changes often, and you would rather edit a paragraph than rebuild a flowchart.
- The team already uses tools it can connect to, so nothing has to be migrated.
If only the first is true, a single AI step inside a workflow tool is usually enough. If none are true, it is probably not an AI job at all.
What a setup looks like
A working Claude Code setup for a business has a recognisable shape.
- One folder, under the business’s own account.
- An instruction file that reads like an onboarding document for a sharp new colleague.
- One skill per repeated job, each short enough to read in two minutes.
- Connections to the record-keeping tools the team already uses, with the narrowest permissions that work.
- A log of what ran and what it produced, in a place a person actually looks.
- A review step wherever output leaves the business.
- Schedules only for the parts that have proven boring.
Who maintains it splits cleanly. Changing what it does is editing plain text, and the owner or an operations lead can do it. Changing how it is wired, permissions, connections, schedules, is engineering work and should be treated like any other change to a system people rely on.
Cost is usage-based rather than per seat, and it becomes predictable once the routine jobs are known. We do not publish numbers because they depend entirely on volume.
What goes wrong
The same five mistakes account for most of the systems we have been asked to rescue.
Treating it like a person who needs no supervision. Instructions so vague that every run produces something different. No record of what it did, so nobody can check. Giving it write access to systems of record before it has earned it. And building a fixed workflow out of it because the AI was the exciting part.
Each one is avoidable at the design stage, which is why we spend the first week or two learning how the work actually runs before anything gets built.
Where to start
Pick one job that is heavy on documents, that a person currently does by hand, and that a person will still review at the end. Write down how it is done today, step by step, including the exceptions. That document is the instruction file in draft form, and it tells you within an hour whether the job is a fit.
If you would rather have us look at the whole operation first, that is what Learn & Recommend is for: one to two weeks, a map of how the work runs, and a recommendation that may or may not include Claude Code.