Stop Telling Your AI to Just Build It. Map It First.

Most business owners open ChatGPT, type their whole idea into one paragraph, and wonder why the plan that comes back sounds finished and falls apart the moment they try to use it.

Tom runs a four-van plumbing company outside Manchester. Good tradespeople, terrible on the phone. Missed calls become missed jobs, and missed jobs become a competitor's job.

Tom decided to fix it with AI. He opened ChatGPT and typed one paragraph: build me an agent that answers my phones, books jobs, and updates my calendar. The output was a tidy diagram and some generic advice about call routing.

He tried again with more detail. Same result. Confident-sounding, wrong in the details, useless the moment he tried to actually build it.

Tom was not doing anything wrong. He was doing what almost everyone does with AI right now: asking it to decide before anyone had actually decided anything.

You Are Asking the Agent to Decide Before You Have Decided

Here is how it goes for most owners today. You have a big idea. You type it into ChatGPT or Claude in one go. Buried inside that one paragraph are ten or fifteen small decisions the AI cannot possibly know: what happens when the caller is an existing customer versus a new one, what happens after hours, who gets called when the agent is stuck, how a price gets read back, what counts as an emergency that skips the queue.

The model does not ask you those questions. It picks reasonable-sounding defaults for all of them, in silence, and hands you something that looks finished. You build it anyway, because it sounds so sure of itself. It breaks on the fourth phone call, on the exact decision nobody made.

The problem is not the model. The problem is the order of operations.

How Do You Plan an AI Agent Project Before You Build It?

You map the open decisions before anyone touches a build session. A developer named Matt Pocock, best known for the TypeScript course Total TypeScript, built a workflow for exactly this gap and shared it in late July 2026. He calls it Wayfinder, and it has become one of the most talked-about AI workflows among the engineers who build agents for a living, racking up over a thousand likes and 51,000 views on a single post about it within hours of going out.

Here is the difference in practice. Ask a chatbot to build you an appointment agent, and it answers once, guesses at every gap, and hands you a finished-looking thing. Ask an agent running Wayfinder, and it does the opposite. It refuses to guess. It interviews you, one question at a time, until every open decision either has an answer or is written down explicitly as a question still to be answered. Nothing gets built until that list is empty.

Pocock's own name for the unresolved part is "fog of war," borrowed from strategy games: the parts of the project you cannot see clearly yet. The unresolved decisions are the "frontier," the edge of what you can tackle right now. The AI's job, in his words, is "to find the way, resolve the open decisions, not to build the thing." Building comes later, once the fog has cleared.

Why This Is Suddenly Everywhere

This did not come from a big lab. It came from one developer publishing his own working method and letting other builders steal it, which is exactly the point: you do not need a research team to adopt this, you need a tracker and twenty minutes.

Within two days of Pocock posting an update to Wayfinder, another builder had already forked the idea, adding what he called "its own harness, with an interactive star map you start tickets off of." That is the tell that a method has hit a nerve: people do not fork things they think are a gimmick.

In a live demo, Pocock ran Wayfinder against a genuinely large feature, a TikTok-style publishing tool bolted onto an existing app, and resolved eight of nine planning decisions inside a single hour-long session, entirely through the AI asking him questions, before a single line of the feature was written.

Isn't This Just Project Planning?

Yes. That is exactly the point, and exactly why it works.

AI has made building things fast enough that the bottleneck moved. It used to be the coding. Now it is the thinking. Wayfinder is not a new kind of intelligence, it is an old discipline with a patient interviewer attached: define what done looks like, list what you do not know, resolve the unknowns one at a time before you commit to a build. Project managers have run versions of this with sticky notes for decades. What is new is having an AI organized enough to keep score, patient enough to ask the tenth follow-up question, and honest enough to say "I don't know yet, let's put that on the map" instead of guessing.

The Method, Adapted for a Business Instead of a Codebase

You do not need GitHub, GitLab, or a single technical bone in your body to run this. You need a shared place to write things down and an AI you are willing to let interview you before it builds anything.

1. Name the destination. Write one or two sentences describing what done looks like. Not "an AI phone agent," but "a caller gets their price read back and sees the job on my calendar within two minutes, with no human touching the call unless it's flagged as an emergency."
2. Open a tracker. Swap the issue tracker for a Trello board, a Notion database, an Airtable base, or even a shared Google Doc with one row per decision. Title it "[Project] Map" and keep it visible to anyone else involved.
3. Have the AI grill you, breadth-first. Tell it plainly: "Before you build anything, ask me every open question you can think of about this project, one at a time, and wait for my answer before moving to the next." Expect questions like: what happens with an existing customer versus a stranger, what counts as after-hours, who gets alerted if the agent is stuck, how price exceptions get handled, who reviews the transcript before this goes live.
4. Sort each decision into two piles. Some only you can answer, a policy call, a judgment call about your business. Others the AI can go find out on its own, a fact-finding task like checking what your booking software's API actually allows. Wayfinder calls these human-in-the-loop and away-from-keyboard decisions. You do not need the jargon, just the split, because it stops you from being interrupted for things the AI can check itself.
5. Resolve one decision at a time. Log the answer next to its row, and let new questions surface as old ones close. This is deliberately slow at first. That slowness is the fog clearing. If the interview turns up almost nothing to decide, the project was small enough to skip the map entirely and build straight away, that is a valid outcome too.
6. Only build once the tracker is clear. When every row has an answer, take the finished map to Claude, ChatGPT, or whichever agent will do the actual work, and say: build exactly this.

What You Have After This

A week in, you have a finished map. Every open decision about the project is named, most are already resolved, and the ones that are not have an owner and a next step instead of sitting in your head unspoken.

A month in, you have a locked plan handed to an actual build session, and usually a first small piece already live, an after-hours voicemail replaced by an agent that triages instead of just recording, while the harder pieces are still being decided properly.

Six months in, you have the full project live, but the bigger win is quieter: you now have a habit for the next one. The next automation you attempt does not start with a hopeful paragraph typed into a chat window. It starts with a map.

Same business. Same tools. A completely different way of deciding what to build next.