Definition first. AI second. That order is the whole method.

Why this exists

AI has made the first draft nearly free. Type a prompt, get something that looks like a finished feature, ship it. The part that's still hard — and still the part that actually matters — is knowing precisely what the thing was supposed to do in the first place, for whom, and how you'd know if it stopped working. Skip that step and you get software that demos well and breaks quietly in production. PromptCraft exists to not skip that step.

How we work

Every engagement starts with definition, not a prompt: what the system needs to do, who it's for, what "working" actually means, and the edge cases nobody wrote down. Only once that's written down do we build — with AI, deliberately and reviewed like production code rather than accepted on the first output — and put together a small evaluation set so every future change can be measured instead of guessed at.

Everything we hand back is yours: the spec, the prompts or code, the eval set, and the reasoning behind each decision. No proprietary framework, no lock-in. See the full breakdown on what we do.

What we're good for

• An AI feature that demoed brilliantly and now behaves unpredictably in front of real users
• A product idea that needs AI embedded properly from the start, not bolted on after
• An internal agent that needs to call tools reliably, not occasionally
• A RAG pipeline that keeps citing the wrong document
• A team that wants to build this discipline in-house, with proper training

What we won't do

We won't skip straight to a prompt without first agreeing what "done" means, and we won't dress up a demo as a finished product. If the honest answer is "you need a different model", "this needs a human in the loop", or "the problem isn't defined well enough to build yet", we'll say so.

Frequently asked questions

What does PromptCraft actually do?

We define the problem properly — what a system needs to do, for whom, and how you'll know it's working — then build it with AI: prompts, agents, or application code, whichever the problem calls for. Then we evaluate it before it ships.

How is this different from just prompting ChatGPT or Claude myself?

Nothing stops you prompting your way to something that looks finished in an afternoon. The part that's usually missing is a written spec of what "working" means and an evaluation set to prove it — without those, every future change is a guess. That's the part we add.

Do you write prompts, or do you build software?

Both, depending on what the problem needs — prompts and agents where that's the right tool, application code where it isn't. The method is the same either way: define first, then build with AI.

Is PromptCraft tied to one AI model?

No. Work spans GPT-4.x/o-series, Claude, Gemini, Llama and other open-weight models — whichever fits the job and the budget.

Where is PromptCraft based?

The UK. PromptCraft is a trading style of Actuality One Limited, a company registered in Scotland.

How do we start a project?

Get in touch and tell us what the system is supposed to do and where it's falling short — we'll tell you honestly whether we can help.