On Friday, the operations manager at a 40-person accounting firm asked Claude Code to build a tracker for client document requests. By Sunday night it worked: a list of requests, a status for each one, and a page to upload files. On Monday morning, the managing partner asked one question: “How do our 300 clients log in, and how do we make sure each one only sees their own requests?”
That question is the gap between generating code and building a business app. Claude Code is very good at the first part. This guide covers the Claude Code best practices that get you the most out of it, and what it takes to turn what it builds into something your team, clients, and partners can rely on every day.
What can Claude Code build for a business?
Claude Code is Anthropic’s AI coding tool. It runs in your terminal, code editor, or desktop app, reads the files in your project, and writes, runs, and changes code based on what you ask it in plain English.
For a growing business, that opens up a lot: a job tracker for field crews, an intake form that feeds a database, a dashboard that pulls numbers from three spreadsheets, or a script that cleans up your invoice exports every Monday. Things that used to need a developer for weeks can now be drafted in an afternoon.
What Claude Code produces is code. Whether that code becomes an app your business can run on depends on everything around it, and on how you work with Claude Code to build it. If you’re weighing the cost side of this, our guide to Claude pricing covers what building and maintaining your own tools costs beyond the subscription.
What’s the difference between generating code and building a business app?
Generated code answers “does it work when I try it?” A business app has to answer different questions: who can log in, what each person sees, where the data lives, what happens when something goes wrong at 7am, and who fixes it next month.
Here’s what sits between the two.
| What a business app needs | What it means in practice | Who owns it if you build from scratch |
|---|---|---|
| Hosting | The app runs on a server, not your laptop, and stays online | You: pick a provider, set it up, pay for it, watch it |
| A database and backups | Records are stored safely and can be restored if something breaks | You: set up the database and test that backups actually restore |
| Logins | Team, clients, and partners sign in, reset passwords, and lose access when they leave | You: build or add a login service and keep it secure |
| Permissions | Each person sees only their records and fields, such as a client seeing only their own projects | You: write the rules into code and re-test them after every change |
| Your process and its exceptions | Approvals, handoffs, and the cases that don’t follow the normal path | You: describe them clearly and update the code when they change |
| Updates and fixes | New fields, new steps, and bug fixes every week | Whoever built it, or whoever inherits the code |
Claude Code can help with every row of that table. The difference is that each row is something you, or someone on your team, now owns and has to keep working.
What are the Claude Code best practices for building business apps?
Anthropic’s own Claude Code best practices come down to three ideas: tell Claude what it can’t see, give it a check to run, and enforce the rules that can’t have exceptions. Here’s how those apply when the thing you’re building is a business app.
1. Describe the process before you describe the screen
Claude Code reads your files very well. What it can’t read is how your business works: that a tax return needs a partner’s sign-off before it goes to the client, or that a subcontractor can’t close a job until the photos are uploaded.
If you’re not sure how to write that down, ask Claude Code to interview you. Anthropic suggests a prompt like “interview me until you could write the spec, then write it to SPEC.md.” Answer its questions about roles, steps, and exceptions, then start a fresh session that builds from that file.
2. Write specific prompts
Vague prompts get vague apps. Name the people, the steps, and what “done” looks like.
- Instead of: “build a client portal.”
- Try: “build a page where each client sees only their own document requests, with status, due date, and an upload button. Partners see all requests for their clients. Clients can’t see internal notes.”
When you want the fuller version, say so. Anthropic notes that Claude tends to build the smallest thing that satisfies a request, and that adding “be ambitious” with the specifics you expect (empty states, error messages, the full save flow) gets you more.
3. Keep your CLAUDE.md file short and useful
CLAUDE.md is a file Claude Code reads at the start of every session. Use the /init command to create a first version, then cut it down. Keep only what Claude can’t work out from the code: the command that runs your tests, decisions your team made, and anything you’ve had to correct twice. Anthropic recommends staying under 200 lines.
4. Give Claude Code a check to run
Claude stops when the work looks done. If it has a test, a build command, or a screenshot to compare against, it can check its own work and fix what broke before handing it back. Ask it for evidence: the command it ran and the output, not just “all tests pass.” After it finishes, ask Claude to run the app and look at the result.
5. Plan first for anything bigger than a small change
For a new feature, switch to plan mode (Shift+Tab) so Claude Code explores and proposes an approach before writing code. It’s much cheaper to change a plan than to undo a day of work. If it heads the wrong way mid-task, press Escape and use /rewind to go back to a checkpoint.
6. Build permissions first, and test as every role
This is where business apps go wrong most often. AI-written code isn’t automatically secure: in Veracode’s 2025 GenAI Code Security Report, which tested more than 100 AI models on 80 coding tasks, the generated code introduced security flaws in 45% of cases.
So decide who sees what before you build any screens, write it into your spec, and ask Claude Code to write tests that log in as each role: a client, a subcontractor, a manager. Each test should check that one client can’t open another client’s records. Our guide to client portal permissions walks through what to plan for.
7. Use realistic test data, not real client data
Ask Claude Code to create sample records that look like your real ones, with the same messy cases: a job with three subcontractors, a client with two offices, an invoice that was partly paid. You’ll find the gaps without putting client data on a laptop or in a test database.
8. Get a second opinion on the code
The session that wrote the code is the worst one to review it. The /code-review command reviews your changes in a fresh session that never saw your conversation. Ask it to check the work against your SPEC.md and to skip style comments, so you get the gaps that matter.
That second check matters because developers themselves are cautious: in the 2025 Stack Overflow Developer Survey of more than 49,000 developers, 46% said they don’t trust the accuracy of AI tools’ output.
9. Connect the tools Claude Code needs
Some of what Claude needs lives in other systems, like your project database or your issue tracker. Claude Code connects to them through MCP, a standard way for AI tools to talk to other software. That lets it read real field names and structures instead of guessing. If your data is in Noloco, you can bring Claude Code into your Noloco app the same way.
What goes wrong when generated code meets real clients and partners?
Most problems don’t show up on the day you build. They show up when more people start using the app.
- Logins. Clients forget passwords, staff leave, a subcontractor needs access for two weeks. Someone has to build and maintain all of that, and keep it secure.
- Permissions that drift. Every new page or field is a new chance to show the wrong person the wrong data. Rules written into code have to be re-tested every time the code changes.
- Exceptions. Real operations are full of them: a job paused for a permit, a client on a custom payment plan. Each one is a code change.
- Hosting and backups. The app has to run somewhere other than your laptop, stay online, and be recoverable if something breaks.
- The builder leaves. If one person built it with Claude Code, the business now depends on someone being able to understand and own that code.
None of these are reasons not to use Claude Code. They’re reasons to decide which parts you want to build and own, and which parts you’d rather have handled for you. For a closer look at this point, read when ChatGPT isn’t enough for building business software.
How do you decide what to build from scratch and what to build on a platform?
A simple test: if the app is internal, used by a handful of people, and someone technical will own it, building from scratch with Claude Code can be a great fit. If clients, subcontractors, or partners will log in, and the process changes often, building every layer yourself means owning logins, permissions, and hosting forever.
There’s a middle path: use Claude Code to build on a platform that already handles those layers.
| What you need | Claude Code, from scratch | Claude Code on Noloco |
|---|---|---|
| Best fit | Internal tools and scripts with a technical owner | Portals that team, clients, and partners log into every day |
| Logins for clients and partners | You build them and keep them secure | Built in |
| Who sees what | Written into code; re-test after every change | Set by role, record, and field |
| Hosting and backups | You set up and pay for each service | Included |
| Where your data lives | A database you set up | Noloco Tables, Airtable, or Google Sheets |
| Making changes | Edit, test, and redeploy the code | Ask Claude Code or edit in the builder; changes to pages, forms and workflows stay as drafts until published |
| Building without a developer | Possible for prototypes, risky for client-facing apps | Yes, with Nola. See how Nola builds it |
How can Claude Code build on Noloco instead of from scratch?
Noloco is the AI portal builder for teams running complex operations. Your team, clients, and partners each log into the same portal and see only what their role allows. And you can build it from Claude Code.
Connect Claude Code to your Noloco app. The Noloco MCP server connects with a one-time sign-in, with no API keys to store. You choose which apps Claude Code can use, and whether it can only read or also make changes.
Ask for what your process needs. For example: “turn my Jobs and Clients tables into a client portal,” “add a form for subcontractors to upload site photos,” or “set permissions so each client sees only their own projects.” Claude Code builds tables, pages, forms, workflows, and permissions in Noloco. Changes to pages, forms and workflows stay as drafts until someone publishes them, and Claude Code checks with you before publishing or deleting anything.
Logins, hosting, and permissions are already there. You’re not writing login code or setting up servers. Permissions work by role, record, and field, so a subcontractor can see a job address without seeing your margin.
Keep your data where it is. Noloco runs on Airtable, Google Sheets, or its own built-in database.
No developer? Use Nola. Nola, Noloco’s own AI builder, builds portals, dashboards, and workspaces from a plain-language description, so the person who owns the process can build it directly.
What it costs. The MCP connection is included on every Noloco plan. The Free plan includes unlimited team seats and 50 active clients. The Build plan includes 10 team seats (then $5 per extra activated seat) and 1,000 active clients. See Noloco pricing for details.
Final thoughts
The best way to get the most out of Claude Code is to treat it like a very fast, very capable builder who knows nothing about your business. Give it your process in writing, give it a way to check its work, and have someone else review what it built.
Then be honest about what you’re making. A script or an internal tool can live as code. An app your clients and partners log into every day needs logins, permissions, hosting, and an owner. For that kind of app, using Claude Code to build on a platform that already handles those parts usually gets you further, faster, with less to maintain.
Next step: see what building your own tools really costs in our guide to Claude pricing.
Frequently asked questions
What are the most important Claude Code best practices?
Tell Claude what it can't learn from your files (your process, decisions, and rules), give it a test or check it can run, ask for evidence that the check passed, and have the code reviewed in a fresh session. Use plan mode for larger changes.
Can a non-developer use Claude Code to build a business app?
Yes, for prototypes and internal tools. For an app that clients or partners rely on, someone still needs to handle hosting, secure logins, permissions, backups, and ongoing fixes, either a developer or a platform that takes care of those parts.
What should go in a CLAUDE.md file?
Only what Claude can't work out from the code: the commands to run tests and builds, decisions your team has made, parts of the code that are misleading, and anything you've had to correct more than once. Keep it short.
Is code written by Claude Code secure?
It can be, but don't assume it. Independent research has found security flaws in a large share of AI-generated code. Plan permissions up front, write tests that check each user role, and have the code reviewed before real users touch it.
How do I stop Claude Code from building the wrong thing?
Write a short spec first, or ask Claude Code to interview you and write one. Use plan mode so you can approve the approach before it writes code, and stop it early if it heads in the wrong direction.
Can Claude Code connect to my existing business data?
Yes. It can work with files and command-line tools directly, and it connects to other systems through MCP servers, which let it read and, if you allow it, change data in those systems.
Related resources
Written by
Marta Prunés
Content Marketing Manager at Noloco