Your AI assistant gets it right and wrong in the same answer. It forgets your rules two messages in, it invents facts about your company, it writes in a voice that is not yours, and you keep pasting the same corrections by hand.
That is often not a model problem. It is usually a missing or very thin system prompt.
We write one for you, properly, for the model you already use.
WHAT YOU GET
1 custom system prompt, 400 to 900 words, written for the model you name in your order: ChatGPT or a custom GPT, Claude, Gemini, a local model such as Llama or Qwen, or an agent framework.
1 PDF document, about 6 pages, with the full prompt plus a plain explanation of every section, so you can edit it yourself later.
1 plain text file with the prompt only, ready to paste into your system field or config file. Nothing to clean up.
3 example conversations, written out in full: one everyday task, one awkward edge case, and one attempt to talk the assistant out of its rules, so you can see how the guardrails behave under pressure.
1 guardrail list: what your assistant refuses, what it asks you about instead of guessing, and what it must never claim.
1 revision round, included, within 7 days of delivery.
Prompt language: English or German, your choice.
HOW IT WORKS
1. The order form asks you nine questions, about ten minutes of your time: role, audience, tone, hard rules, output format, and three real examples of the work you want done. You answer them in the order box.
2. We draft the prompt, then run it in test conversations using your own examples, then fix what breaks. What you receive is the version that came through that test, not a first draft.
3. Where we cannot reach the exact model you named, for example a self hosted or fine tuned model on your own machine, we test on the closest model we do run and write that plainly in the PDF. You always know what was tested and on what.
4. AI drafts and runs the test conversations, a human reads every line, corrects it and approves it before it goes out. Human with machine. We say it openly, because you are paying for the result, not for a story about how it was made.
WHAT WE NEED FROM YOU
1. Which model and which interface you use.
2. What the assistant is for, in one or two sentences.
3. Who it talks to: customers, your team, or you alone.
4. The tone you want, or a short sample of writing you like.
5. The rules it must never break, for example never give legal advice, never promise a delivery date, never switch language.
6. The output format you want back: bullet points, email, JSON, a fixed template.
7. Three real examples of the task, good or bad, so we can test against reality.
8. Any company facts, product names or wording it has to get right.
9. Prompt language, English or German.
Send it all in one message in the order box. If anything is missing, we ask once, in writing, and agree a new delivery date with you through this order. Please do not send passwords, API keys or account logins, none of that is needed.
DELIVERY
3 days from your order, when your answers are in the order box with it.
Need it faster? Add the 24 hour option at checkout.
Also at checkout: a second prompt variant, an agent version for assistants that call tools, five extra test conversations, and two extra revision rounds. Every one of them is a one off add on to this order.
WHAT IS GUARANTEED
The delivery itself. The files listed above, in the formats described, on time, with one revision round if the prompt does not behave the way we agreed.
We put no number on the outcome. No ranking, no traffic figure, no percentage. Language models are not deterministic and the vendors keep updating them, so an identical reply every single time is not something anyone can promise, and we will not pretend otherwise.
This is a new listing, so it is written down to the file count and the delivery day. Hold us to it.
Order, answer the nine questions in the order box, and your prompt is with you in three days.
