How I work with AI agents: from an idea to something useful

My AI toolkit changes. The goal stays the same: make something useful, put it in front of the people it is meant to help, and learn what to improve.

I work across product, design, development, teaching, and marketing. AI agents help me move between those parts of the work. I might start with a rough idea, work through the experience, build a page or tool, create supporting content, and then come back to the customer journey.

The interesting part is how those pieces connect. Here is the process I use, and a small version you can borrow without buying every tool I mention.

1. I start with the result someone needs

“Make a website” leaves a lot undecided. “Help a small-business owner understand this service and request the right kind of help” gives the work direction.

I bring my product, UX, and marketing experience into that first decision. Who is this for? What are they trying to accomplish? What might confuse them? What would make the next step feel easier?

The result might be a clearer offer, a working prototype, a lesson someone can complete, or an inquiry process that makes sense. Naming it helps me decide what to ask the agent to do next.

2. I give the agent the project, not just a prompt

Before changing a site, I want the agent to look at what is there. Before writing an article, I want it to understand the audience, voice, existing material, and destination. Before recommending an improvement, I want it to separate what we know from what we are guessing.

Useful context includes the current page or files, approved facts, brand examples, constraints, and the definition of finished. It also includes decisions we have already made. Repeating an old idea because the assistant missed that decision creates more work.

For a concrete example, see the earlier walkthrough of turning one rough brief into a published article. This broader process applies beyond writing.

3. I organize the work by roles

I use assistants for different kinds of work: finding information, exploring options, building, checking, and preparing the supporting content. A role can be another pass in the same conversation; it does not always require another agent or another subscription.

• Research: gather relevant sources and identify open questions.
• Product and design: shape the journey, structure, copy, and visual direction.
• Build: implement a focused change and explain what changed.
• Review: check the result against the brief and inspect the actual experience.
• Marketing: adapt the useful idea into content for the audience and channel.

Tasks can move alongside one another when they are independent. When one task depends on another, I want a clear handoff: the approved brief before the build, the checked facts before the final copy, and the working destination before promotion.

4. The tools have jobs in that process

My toolkit spans ChatGPT and Claude for thinking, drafting, and agent-assisted work; Figma and Lovart for design exploration; coding agents for implementation; and Higgsfield, HeyGen, and ElevenLabs for visual, video, and voice projects. Railway is part of my hosting setup, and Supabase is part of the data layer for some projects.

On the marketing and operations side, the work can involve n8n or Make, plus connections to tools such as HubSpot, MailerLite, Resend, and Brevo. Browser access and app connections help when a task needs information or an action in another system.

That is a picture of the tools across my work, not a shopping list. I do not use every one on every project, and a tool being available does not mean every connection or automation is configured. Start with the tools you already have and one result you can judge.

5. I stay involved in the work

I ask for revisions. I change the hierarchy. I look at whether an image helps explain the idea. I want the wording to sound like me and the experience to make sense to the person using it.

A page can look finished and still have a confusing button. A paragraph can read beautifully and contain a claim we cannot support. Code can be written while the deployment is still incomplete. Those differences matter.

I use a simple review question: can the intended person actually use this? That leads to checking the mobile layout, links, instructions, form behavior, and whatever else the task requires. A second AI review can help, but I still need evidence from the result.

6. I close the loop after “done”

For a blog post, finished means a working article that people can find from the site. For a build, it means the relevant behavior works in the environment where it will be used. For an automation, it means we know what starts it, what it does, and who checks an exception.

Then comes learning. Publication is not proof of readership. A working tool is not proof that people want it. I want to understand what people use, where they get stuck, and which next improvement is worth the effort.

Try a small version this week

Choose one task you already need to finish. Keep it small enough to review in one sitting: improve one service section, turn approved notes into a lesson outline, or create a useful article from material you know well.

Help me finish [specific deliverable] for [audience]. The result should help them [next action]. First review [approved sources or files]. Keep [constraints and existing decisions]. Identify missing facts. Propose the steps, then prepare the work for review. Show what changed, what you checked, what remains unfinished, and where the result is saved.

Keep a note of preparation, review, and correction time as well as the first draft. Use the workflow run log if you want a place to record your attempts. The goal is to find a process you can use again.

You do not need to become an expert in every AI tool before you start. Bring what you already know, choose one useful job, and give yourself room to learn while doing it.

Useful AI beats hype. Real prompts. Real projects. Real results.

Want the team version? Read how we organize AI-assisted client delivery. Building a product? Explore the product-building version of this process.

Follow Ashley’s founder journal · Try the build-along challenge · Follow the field notes