I asked ChatGPT to write my project brief. Here's what it quietly got wrong.

Four years ago, the advice for writing a project brief was “use this template.” Two years ago it was “use this prompt.” Today you can describe your app in one sentence and get back a polished, professional specification in about thirty seconds — sections, workflow, requirements, success metrics, the works.

So the brief is a solved problem now, right?

Not quite. I ran the experiment on a realistic project, and here's the surprising part: the output was good. Restrained, sensible, professional — which is exactly what makes its failures dangerous. They're not the obvious over-engineering everyone warns about. They're quiet. You notice them when you get the bill, or when the core feature turns out to be impossible.

The experiment

I gave a chatbot one honest sentence, the kind a real founder would type:

Thirty seconds later, this came back (condensed):

Honestly? Not bad. It structured the workflow correctly and reads like a professional spec — authoritative enough that nobody thinks to question it. But notice what didn't happen: it never asked me a single question. Not how many users, not whether inspectors belong to different companies, not which countries or languages, not whether anyone approves an inspection before it's final. It just decided all of it, silently, and handed me a document that looks like the decisions were mine.

That's the thing about a generator: it never asks; it fills. And a filled gap looks identical to a decided one — until it costs you.

The three words that could sink the whole project

Read that requirements section again: “integration with public land registry APIs or data sources where available.”

Where available. The entire app lives or dies on that clause. Public land registries are a mess — fragmented by jurisdiction, frequently with no public API at all, sometimes accessible only by manual lookup or scraping a government portal that changes without notice. Whether this integration is a weekend of work or a six-month wall of bureaucracy is the single most important unknown in the project — and ChatGPT reduced it to a breezy subordinate clause and moved on to describe the camera workflow in loving detail.

That's the pattern in miniature: confident and thorough about the easy parts, three-words- and-a-shrug about the existential one. A generator optimizes for a document that looks complete. “This might be impossible, we need to check first” doesn't fill a section nicely, so it gets compressed into “where available” and buried.

The questions a human would have asked first

Hand that same sentence to a competent engineer and the first thing they do is ask questions — because they know the answers change everything. Here's a fraction of what ChatGPT decided for me instead:

How many users, and how many kinds?

One inspector, or thousands? A single role, or inspectors plus supervisors plus admins? The answer reshapes the auth model, the database, and the price.

Is it multi-company?

Do inspectors belong to different firms — subsidiaries, franchises, competitors who must never see each other's data? Multi-tenancy is an architectural decision you make on day one or pay dearly to retrofit. The brief silently assumed a single flat pool of users.

Which countries, languages, and conventions?

Property law, address formats, registry systems, and date conventions differ by country. “The public land registry” isn't one thing — it's a different thing in every market. The brief assumed one.

Each of these is a genuine fork with real cost on each branch. A human asks because they can't responsibly guess. A generator guesses because a question mark doesn't look like progress — and the guess arrives dressed as a decision you supposedly made.

The silent decision that could gut the core feature

Look at “document capture with automatic image organization.” The AI decided your documents are photos to file — not text to read. There's no OCR, no extraction, anywhere in the brief.

But the app's whole point is to compare the property against the registry. If the parcel number, owner name, and boundaries live on a photographed deed, you need to pull that text off the image to compare it — that's OCR (Textract, Tesseract, whatever), and it's a real, non-trivial piece of work. The AI silently assumed you didn't need it. Maybe you don't. But that's a scope decision worth thousands of euros, and it was made for you, invisibly, in the space between “photograph” and “organize.” You'd discover it the day someone asks why the app makes them retype every parcel number by hand.

The cost multipliers you'd never spot

It's restrained, but not immune. Several things crept in that a validation MVP doesn't need — and two of them are the kind of quiet cost multipliers a non-technical founder has no way to price.

Offline capability with later sync

This one sounds prudent — inspectors work at remote sites, so of course it should work offline. Except offline-first with conflict resolution is one of the most expensive things you can put in an app; it can multiply the build cost several times over. Nobody asked for it, and the founder can't judge what they agreed to, because it's phrased as a sensible requirement rather than “this could double your budget — do you need it on day one?”

A hidden state machine

The brief mentions inspections that get “submitted” and “flagged.” Innocent words — but they smuggle in a question nobody asked: is there an approval workflow, a supervisor who reviews and sends it back? If so, every inspection now has states (draft, submitted, under review, approved, rejected), transitions, per-state permissions, and a role that didn't exist a sentence ago. That's the difference between a form that saves data and a stateful workflow engine — hidden inside two verbs.

An audience that quietly quadrupled

My sentence said “independent inspectors.” The brief's target users became inspectors, surveyors, lenders, and legal firms — each a different workflow and sales motion. The AI expanded the market because a bigger list looks more thorough, and just like that your focused MVP has four masters.

What it still didn't force you to decide

A brief's actual job isn't to look complete. It's to make you decide the things that are expensive to change later. This one — good as it looks — skipped every hard fork:

What happens when two inspectors log the same property? When an inspector deletes their account, what happens to their filed reports and photos? What happens when the registry lookup fails or doesn't exist for that region — does the inspection proceed, block, or fall back to manual? None of these have a default answer. The generator skipped them precisely because they're unresolved — there's nothing confident to write, so it wrote nothing.

The lesson — and it isn't “AI is bad”

I use these tools every day; they're remarkable, and generating a first-draft brief is a fine use of one. The problem isn't capability. It's goal. Ask an AI to write your brief and it optimizes for a finished-looking document — confident in every section, hedging the scary parts into subordinate clauses, deciding the ambiguous ones silently. A brief that papers over ambiguity is worse than no brief, because it hides the decisions instead of surfacing them.

So the AI tool you actually want isn't the one that writes your brief — it's the one that reads it and tells you what you haven't decided: the buried “where available,” the silent OCR assumption, the questions it never asked. Something built to ask instead of assume.

That's why FSGen starts with one. A fixed price only works if the scope is genuinely nailed down first — so you run your brief through a free feasibility check that flags what's still open, surfaces the assumptions a generator buried, and tells you whether the project is defined enough to build on a fixed timeline at all. Sometimes the honest answer is “not yet — decide these three things first.” The two-week build only makes sense once it passes.