How to build with AI: from an idea to a working system

Let's say you have an idea for an app. Or you want to create a system for your company that automates and delegates to AI everything that can be delegated. You know what needs to change and how you'd like it to work. You have a vision: something that doesn't exist yet and could make your life easier or generate income.

Today you can turn that idea into a program, an app, a game, a book, a website, a blog or a business system. With a normal computer and an AI subscription. The leap we've made in just a few months is extraordinary: you direct the building process by explaining what you want to achieve.

You can now try Codex on the Free and Go plans, although it is not reaching every account at the same time. For 20 dollars a month, you get ChatGPT Plus or Claude Pro. For serious work, 100 or 200: ChatGPT Pro or Claude Max. I'd start with a single subscription and upgrade when the work calls for it.

Chatting with Claude or ChatGPT on your phone is one thing; making the leap to creating something that works is another. You may have heard of vibe coding: building software by talking to an AI, testing what it does and asking for changes. And not just software: documents, presentations and any other content too.

I can vouch for it: it works. I've already explained how I build this blog with AI and how I prepared a professional presentation. Here I lay out the journey so you can save time and money when you start.

This article opens a series: each step will have its own instalment, with the detail needed to put it into practice. Today, the whole picture: from giving AI tools to organising several bots and deciding what they can do on their own.

1. Install AI on your computer and give it permission to program

To use Codex, download the ChatGPT desktop app from its official website. Sign in with your account, choose Codex and open a folder for your project. In the permissions control below the text box, start with Ask for approval: it edits within that folder without asking permission for each change and asks your permission to go beyond it.

If you prefer Claude, install its desktop app and sign in with your Pro or Max account. Open the Code tab and, below the text box, choose Local and select the folder. There are four main modes: Manual (you approve each change and each command), Plan (it reads files and runs exploratory commands to propose a plan without editing source code), Accept edits (it accepts edits and common file operations, such as creating folders or moving files, and asks you about other commands) and Auto (it reviews actions automatically).

With that, your AI is now your programmer: it reads files, writes code and tests it.

2. Set up a repository system

Before you accumulate files called “final”, “the good final” and “this time for real”, ask the AI for a folder per project with a Git repository inside. Git keeps the change history. Each commit is a point you can return to, with an explanation of what you've changed. Save one whenever something works.

Ask like this: “Check whether this project uses Git; if not, set it up and save a first version we can recover.”

A private GitHub repository gives you an external copy and a place to collaborate. But uploading the project is a separate decision: a local commit is still only on your computer.

3. Add Hermes when you need to organise more work

If you want to give your AIs more capabilities, I recommend Hermes Agent, the open-source agent from Nous Research. Follow its official documentation to install it.

With Hermes, you create profiles with their own instructions and skills — reusable working recipes — and delegate tasks to other agents. That's how you organise specialists and distribute assignments. You also schedule recurring tasks (crons) and triggers that start something when something happens.

On top of that, I've built my ticket system: it organises projects, keeps processes going and alerts me when something stops.

4. Measure usage by bot and by project

You need to know what each bot consumes and what it costs to move each project forward. You can have plenty of activity and very little to show for it.

I measure by bot and by project. I want to see where the resources go and what useful work I get in return. Ask the AI to record the usage information each provider gives you, the tasks and their results, and to separate the subscription from any extra spending.

I've already explained what AI assignments really consume. The rule is simple: measure before you multiply the bots.

5. Talk to your bots from your phone

When work carries on without you sitting at the computer, you need a channel for receiving results and giving instructions.

Hermes connects to messaging services such as Telegram. Your phone becomes your command post while the bots work on your computer. Ask it to set this up and check that you can send an assignment, receive the result and correct it.

I step in when needed. Without watching the screen.

6. Start with a bounded environment and a working prototype

Start in a folder on your computer, with test materials, and ask for a prototype that does what you want.

Don't start by asking for a perfect architecture, ready for every imaginable problem, before you've even tested the main function. You'll end up with thousands of lines of code that block progress and make any change difficult.

First, the idea working on a simple foundation. Then, what it really needs.

7. Keep what works and test improvements in parallel

Once you have a version that works, save it and ask for a separate copy for experiments.

Try different models on the same assignment: which solves it best, how long it takes and how much it consumes. If a configuration works for you, clone it and improve one of the two. Keep the other as a reference.

That gives you something to compare against, and somewhere to return to when an improvement makes things worse.

8. Simplify processes and reduce usage

With the system running, look for what a normal program can do, without waking an AI for every step.

When intelligence is needed, choose the right level and give it only the context it needs. Sizing AI to fit the work is also part of building well. A task repeated twenty times a day with ten thousand tokens of context doesn't cost the same as one with a million. The context window can become a token vacuum if every assignment drags the entire history along.

The same result, with less pointless work.

9. Security without paranoia: control what goes out

Before taking the project off your computer, prepare the access controls and approvals.

AI works freely within the environment I've given it. Everything that goes outside comes through me, one approval at a time. Preparing an email isn't sending it. Using up the allowance isn't buying more, renewing a plan or making a payment.

The initial permissions aren't enough: you need technical limits, as I explain in how to limit an AI agent.

Inside my own home, freedom to build, test and improve. I keep the key to the door.

Official sources

Consulted on 25 October 2026:

1. OpenAI: desktop app and first steps, permissions and plans with Codex.
2. Anthropic: getting started with Claude Code on desktop and permission modes.
3. Anthropic: Claude pricing and Max tiers.
4. Nous Research: Hermes Agent documentation and official repository.
5. GitHub: Git and GitHub.