The bot gives different answers each time and sometimes invents facts
We write a system prompt covering role, limits, answer format and what to do when data is missing. We add reference answers and test it against a set of real questions.
Make your LLM answer the way your business needs, every time
We write system prompts and prompt libraries that keep a model on your tone, facts and rules. They sit under your chatbot, AI agent or team workflows.
Quote and schedule after your brief
Who it's for
- Companies whose chatbot answers inconsistently or makes things up
- Teams that use ChatGPT or Claude daily and rewrite prompts from scratch each time
- Owners planning an AI agent for lead handling, support or sales
- Product teams embedding an LLM into their own service
Problems we solve
Everyone on the team prompts differently, results are unpredictable
We build a prompt library for your recurring tasks: emails, product copy, customer replies. Each prompt has variables, an example and a usage note.
A model update broke what worked yesterday
We build an evaluation set: a few dozen typical inputs with expected outputs. After any model change you rerun it and see exactly what shifted.
The model doesn't know our pricing, policies or voice
We restructure your knowledge into a format models read well and define your tone using real examples of your emails and conversations.
What's included
- Task analysis: who the user is, what a good answer looks like, where the limits are
- System prompt: role, tone, restrictions, output format
- Few-shot examples (input and ideal-output pairs that teach the model your style)
- Failure handling: what to say when there is no answer and how to hand off to a human
- Prompt injection defenses (against attempts to make the bot ignore its rules)
- Prompt library with variables and notes on when to use each prompt
- Evaluation set and a version comparison sheet
- Model selection for the task and budget (GPT, Claude, Gemini, open models)
- Structured output (JSON) when answers feed a CRM or code
- Documentation and handover of all files
Choose the scope
Use a package to describe what you need. We agree on the quote and schedule after reviewing your brief and requirements.
01
Starter
One task or one chatbot that needs to be brought under control
- System prompt for one scenario
- 5–10 reference answers
- Evaluation set of 20 questions
- Short guide on editing the prompt
02
Library
A team using AI daily in sales, support or marketing
- Up to 15 prompts for your workflows
- Variables, examples and usage rules
- Evaluation set and comparison of two models
- One-hour team training
03
Product system
An AI agent or LLM feature inside your service or CRM
- Instruction set for several agent roles or steps
- Structured output and prompt injection defenses
- Extended evaluation set and a pilot with a KPI
- Handover to your developers or connection to a Yucod AI agent
How we work
Brief and requirements
You describe the task and current workflow. We clarify requirements, choose a first stage and prepare a quote.
Material collection
You share sample emails, conversations, pricing and rules. We decide what the model should know and what it should not.
First version and tests
We write the prompts and run them against the evaluation set. You see where answers are still weak.
Pilot with a KPI
We run it on a small share of real conversations. The metric is agreed upfront, for example the share of answers sent without edits.
Handover
You get the files, access and documentation. All accounts and API keys are yours under the agreed terms.
Scope and limits
- We don't promise a model will never be wrong. We reduce failures and teach it to say "I don't know".
- We don't replace missing data. If you have no current pricing or policies, a prompt can't invent them.
- We can't guarantee identical quality after a provider changes a model. That's why you get an evaluation set to recheck.
FAQ
How is prompt engineering different from building a chatbot?
A chatbot is the interface, integrations and logic. Prompts define the model's behavior inside it. You can order prompts alone if a bot already exists. If it doesn't, see our AI chatbot service.
What do I need to prepare?
Real conversations or emails, pricing, rules and typical refusals. The more real examples, the more accurate the result. If you don't have them, we collect them together at the start.
Does this work with ChatGPT, Claude and other models?
Yes. The principles are shared, but wording is tuned per model. Comparing two models on your evaluation set is included in the Library package and above.
How do you check quality?
We build a set of typical and difficult questions with expected answers. Every prompt version runs through it and the results go into a sheet. You see the trend, not vague claims.
Can I pay in stages?
Yes. Payment is split into stages: after the brief, after the first version and after handover. All files and access are yours under the agreed terms.
Do you need access to our data?
Only what the task needs. Customer personal data can be replaced with test data. You create API keys in your own account and can revoke them anytime.
Start with your task
Describe your business, the outcome you need and your current workflow. We will clarify the requirements and prepare a quote after your brief.