Claude gave you a full spec for your AI idea. Now what?

Written by the editorial team · 12 November 2026 · Updated12 November 2026

A specification written by Claude or ChatGPT is a good first draft of your intent, not proof that the idea works. Between the spec, a two-hour demo and a system your team uses every week, the real work is on your own data, the cases that break it, a way to measure that it is right, and someone to keep it running. That is the work of a forward deployed engineer, and it is what we do.

You described your idea to Claude or ChatGPT, and a few minutes later you hold an exhaustive specification: user stories, an architecture, a roadmap, sometimes a pricing page. It feels like the hard part is behind you. We know that feeling, and it is real. These tools turn an idea into a structured document better than most people can. What the document cannot tell you is whether the thing it describes will work on your data, with your people, on a Monday morning.

What a generated spec is good for

It puts your intent into words in minutes. It finds what you forgot to ask. It gives you something to hand to a team or a supplier. Read it as a first draft of what you want, not as evidence that it can be built to do what it promises.

What sits between the spec and a product that works

There are three stages, and they are far apart: the spec, the demo and the system people use.

A demo takes an afternoon. It runs on a clean sample, follows the happy path and impresses everyone in the room. A system that people use every week needs the work nobody sees in the demo:

• Real data. Documents spread over five places, names written three different ways, access rights that differ by person.
• The cases that break it. The exceptions the spec never mentioned, which are most of the daily work.
• A way to know it is right. A set of real cases, written before you build and run again after every change. Anthropic puts it this way: “The key to success, as with any LLM features, is measuring performance and iterating on implementations.”
• A person who validates. What the system prepares, someone in your company approves before it leaves.
• Integration and rights. Which tools it reads, who can see what.
• Hosting and data. Where it runs, who can read your documents.
• Upkeep. Models, tools and your own processes change; the system has to follow.

An imagined example

Say you asked for a spec for an agent that answers customer emails. The demo answers the five emails you picked, and it does it well. In the third week of real use, the agent meets an email written in two languages, a forwarded thread of eleven messages, an angry customer, and a refund rule that changed last month and exists only in one colleague’s head. None of this is in the spec. Each case needs a decision: what the agent does, what it refuses, what it hands to a person, and how you will notice next time. That list of decisions is the product.

Is it a product or a good-looking demo?

Seven questions, to ask of any prototype, including yours:

1. Who uses it every week, and what do they stop doing?
2. Which real documents or data did you test it on, rather than a sample?
3. What happens when it is wrong?
4. How do you measure that it is right, and on which cases?
5. Who maintains it in six months?
6. What does it cost to run each month?
7. What does it replace, and did the people concerned agree?

If you cannot answer three of them, you have a demo. That is useful, and it is not yet a product.

Sometimes the simplest solution is the right one

Anthropic’s guidance on building agents starts there: “we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” A spreadsheet, a tool you already pay for or one well-written prompt sometimes does the job. Saying so is part of the work.

Where expertise changes the outcome

Choosing the smallest slice that is worth building. Writing the test cases from your real situations. Connecting to your documents with the right access. Putting it into production and following it afterwards. This is the work of a forward deployed engineer, and doing it well takes experience.

And with us?

We are a forward deployed AI agency. The first call lasts 30 minutes, it is free and carries no commitment. You authorise access to your tools in two clicks, without a password, and two weeks after the call a first agent runs for your team to test. A person in your company validates before anything is sent, and a dedicated contact follows the project. We take on agents, and also machine learning and data projects.

If you are weighing building it yourself with a custom GPT or a Claude Project, see the page on how far a custom GPT or a Claude Project goes.