Prompt engineering

A working prompt almost always consists of the same blocks. Role and framing: who is answering and in what situation. Context: the materials to base the answer on, with an explicit ban on inventing beyond them. Task: what exactly to do, stated as concretely as possible. Format: the structure of the response, and for integrations a strict JSON schema that code can validate afterwards. Constraints: length, language, tone, forbidden topics, and what to do when the data is insufficient. A separate technique is a few input-output examples, which fix the format more reliably than any verbal description.

What separates engineering from phrase-hunting is measurement. A prompt that seems to answer well cannot be improved or maintained: any change may fix some cases and break others, and there is nothing to notice it with. The working practice is simple: a set of thirty to a hundred real queries with reference answers or at least acceptance criteria, an automatic run after every edit, and a recorded share of correct outputs. Without that set, switching model or version becomes a lottery and the quality debate becomes an exchange of screenshots.

The limits of the method deserve honest statement too. A prompt cannot add knowledge the model does not have, which is what RAG exists for. A prompt cannot guarantee arithmetic accuracy, which is what code is for. And a prompt cannot reliably defend against a user rewriting the instructions: production systems need separate input and output checks, a restricted set of available actions and logging. The practical conclusion is that a prompt is an important part of the system rather than the system, and solving a product problem with one long prompt tends to hit a ceiling quickly.

Related terms

• LLMAn LLM is a large language model: a neural network trained on vast amounts of text to predict the next token, which is why it can write, summarise and answer, yet it stores no facts the way a database does and can be confidently wrong.
• RAGRAG is a pattern where a language model first retrieves relevant fragments from your own knowledge base and answers from them, which delivers current data and a source citation without any fine-tuning of the model itself.
• AI agentAn AI agent is a program in which a language model does not merely answer but acts: it receives a goal, chooses its own tools such as search, APIs, a database or sending mail, works through steps in a loop and stops when the task is done or a limit is hit.
• EmbeddingsEmbeddings are numeric vectors a model produces from text, images or products so that items close in meaning end up close in the vector space, which is what makes semantic search, recommendations and duplicate detection possible.

Related services

• AI and ML developmentWe build AI that solves a defined task and pays for itself, not demos for the sake of demos. Classifiers, recommendations, text and document processing, LLM assistants, and model integration into an existing product.
• IT consulting and product auditThe most expensive mistakes in software happen before the first line of code: a misread problem, a stack chosen out of a contractor’s habit, and a specification that does not exist. Consulting exists to settle all of that before the development meter starts running: what to build, out of what, at what cost and in which order. The result is a document, not an opinion on a call.
• Motion design and AI video for adsWe produce short video ads: shots generated with AI models, motion graphics on top, then edit, grade and sound design. You get a 16:9 master plus reframed 9:16 and 1:1 cuts, ready to upload to social feeds, Reels, Shorts, marketplace listings and ad accounts. We are straightforward about where AI delivers in a day and where filming or hand-built animation is still the right answer.

Read more

• An AI Assistant on Your Own Knowledge Base: How RAG Works and What It CostsHow a RAG assistant differs from a button-tree bot and from an LLM with no data, where hallucinations come from and what actually reduces them, what SaaS costs against a custom build, and which month the project breaks even.
• Where AI and ML Actually Pay Off in a Business, and Where They Are Expensive HypeMachine learning pays off where there is a repeated decision, a lot of similar data and tolerance for error. Missing any one of the three means do not. Here are the use cases that work, their realistic accuracy, the running costs and how to run a four to six week pilot.
• AI Video Models in 2026: Sora, Veo, Kling, Runway and Higgsfield by TaskA model selection matrix by task, an honest cost per finished second including the three to five iterations everyone actually needs, notes on access from Russia, and a blunt list of what AI video still gets wrong.