ChatGPT or Claude wrote your AI spec: what breaks in production?

If you run a small or mid-sized business, a spec written by ChatGPT or Claude is a good first draft of what you want, not proof the product will work. What breaks in production, once your team really uses it, is almost never in the document. It is your real files and their flaws, the cases nobody planned for, no way to measure that the output is right, and nobody to sign off and keep it running. The ten questions below are a review checklist for your own spec. If you can't answer several of them, you have a demo, not a product yet.

Why does a generated spec look more finished than it is?

You described your idea to ChatGPT or Claude, and a few minutes later you're holding a full document: user flows, architecture, roadmap. These tools write fast and well, we use them ourselves, and the document is useful. It puts your intent into words, it surfaces what you forgot to ask, and it gives you something to hand to a team or a vendor.

What it can't see is your data. It describes a system that works on clean documents, with simple cases and people who have time. It doesn't know your files live in several places, that one customer is spelled several ways, or that an important rule exists only in a coworker's head.

A generated prototype (with Lovable or Bolt, say, tools that build an app from a description) has the same blind spot: it shows a screen that works on the happy path. That's a great starting point. It isn't yet something your team will rely on every week.

What the spec says, and what production does

Spec sentences read well, but they usually hide something:

  • The spec says "the agent reads company documents." In production those documents sit in several tools, in duplicates and in different versions, and some of them should only be seen by a few people.
  • It says "the system answers customer requests." A customer writes in two languages, forwards a long thread, or asks for an exception nobody anticipated.
  • It says "answers are reliable." It never says which cases you check against, or who rechecks after every change.
  • It says "the user approves the result." In practice: who approves, when, and what happens when that person is out?
  • It says "integrates with existing tools." Real permissions, access rights and file formats only show up once you connect.
  • It says "maintenance TBD." Models, tools and your own processes change. With nobody watching, the system drifts without anyone noticing.

What should you ask your spec?

Ten questions, to ask of the document or of a prototype. Write an answer to each one. Each question also says what happens if it stays blank.

  • 1. Which real documents or data did you try it on? If blank: it works on the sample you picked and fails on everything else, usually in the first week of real use.
  • 2. What are the ugliest cases in this work, written down before you build? If blank: they show up in production one at a time, each discovered in front of a customer or a colleague.
  • 3. How will you know it's right? That takes a set of real cases where you know the correct answer, rerun after every change. If blank: a change fixes one case, breaks another, and nobody notices.
  • 4. Who signs off before anything goes out, and on what? If blank: a mistake leaves the company under your name.
  • 5. What does it do when it doesn't know? If blank: it makes up a plausible answer instead of handing off (see AI hallucinations).
  • 6. Who is allowed to see what, and where is your data hosted? If blank: a confidential file becomes readable by someone who shouldn't see it, or passes through a tool your contracts don't allow.
  • 7. Who uses it every week, and what work does it take off their plate? If blank: the system works and nobody opens it.
  • 8. What does it cost to run each month, counting models, hosting and the time to watch it? If blank: the cost shows up after launch, when stopping is harder.
  • 9. Who keeps it current when the AI models, tools and your processes change? If blank: it drifts, and you hear about it from a complaint.
  • 10. Would the simplest option do: a spreadsheet, a tool you already pay for, one well-written prompt? If blank: you build a system for a problem that didn't need one.

How do you read your answers?

If you answer almost all of them without hesitating, your spec is ready to build from, and you don't need us. If several stay blank, you have a demo. It proves the idea is possible. It doesn't prove it holds up on your data with your team.

Anthropic's guide to building agents says "the key to success, as with any LLM features, is measuring performance and iterating on implementations." It also recommends "finding the simplest solution possible, and only increasing complexity when needed," and extensive testing in sandboxed environments with the right guardrails. Those are questions 3 and 10 on the list.

What if your Lovable or Bolt prototype stalls?

A generated prototype makes a very good living spec: your colleagues react to a screen instead of a long document. It usually stalls where the document does, when it meets your real data, your access rules and your exceptions, and when someone has to host it, monitor it and keep changing it.

Don't throw it away, and don't keep stacking prompts on top. Write down the cases that make it fail, rerun them after every change, then decide what you keep. When those cases go beyond what you want to handle yourself, that's the moment to bring in help.

Can you build it yourself?

Often, yes, especially with developers on the team. Writing the code is rarely the hard part. For one focused task on a few stable documents, a custom GPT or a Claude Project sometimes does the job; the page on how far that goes shows where the line is.

The work changes once the tool has to read your real files and other people depend on it. To see where you stand before deciding, an AI audit of your processes and data is a good first step.

What does a forward deployed agency do between the document and the product?

Engineers work with you, on your data, instead of shipping software and leaving. In practice:

  • pick the smallest slice of the project that is worth building;
  • write the test cases with you from your real situations, the ugly ones included;
  • connect to your documents and tools with the right access, read-only;
  • have someone at your company approve what the agent prepares before anything goes out;
  • put it in production, then follow it, with a dedicated contact (see what happens to an agent after delivery).

And at Saména?

Saména is a forward deployed AI agency. We build custom agents for work that handles data (email, quotes, case files, documents), and we also take on machine learning and data projects. The agents cite their sources, and nothing goes out without a person at your company signing off.

If you have a spec or a prototype, bring it to the first call: a free 45-minute discovery call, no commitment, with Ange or Xisco.

Frequently asked

Is a spec generated by ChatGPT or Claude reliable?
It's a good first draft: it puts your intent into words and surfaces what you forgot to ask. It doesn't tell you whether the system will work on your data with your team. To find out, you have to try it on real documents and real cases.
Why does my prototype work in a demo but not in real use?
A demo runs on a clean sample and follows the happy path. Real use brings your actual files and their flaws, exceptions nobody wrote down, and people who don't behave as planned. Write those cases down and rerun them after every change.
How can I tell whether an AI demo will hold up in production?
Run it on real documents and on cases you wrote before building, the ugly ones included. Rerun the same cases after every change, check what the system does when it doesn't know, and decide who signs off before anything goes out.
Do I need to rewrite the spec before building?
Rarely. It's a starting point. What it needs is what it leaves out: the real data, the cases that break it, a way to measure that it's right, sign-off by a person, hosting and upkeep.
What should I bring to a first call?
Your spec or prototype if you have one, a process that costs your team time, where its documents or data live, and what a good result looks like. Nothing formal.

Bring your spec or your prototype to the free 45-minute discovery call. No commitment.