When Harvard and BCG put 758 consultants to work with generative AI, output quality improved by 40% and tasks were completed 25% faster. But the same study, published in Organization Science in 2026, revealed the other side of the coin: on tasks where AI is unreliable, those who trusted it anyway performed 19 points worse. Knowing how to ask the right things of AI — the much-discussed practice of prompt engineering — matters; knowing when not to trust the response matters even more. This guide separates, with evidence from 2025–2026, the techniques that still work from those that have become mere noise, and explains why training your team in this has stopped being optional: since February 2025 it is a requirement under the European AI Regulation.
What is prompt engineering?
Prompt engineering is the practice of designing the instructions you give to a generative Artificial Intelligence — the prompt — in order to obtain useful, accurate results in the format you need. It covers what you ask for, how you structure it, what context and examples you provide, and what constraints you set. It is not programming: it is communicating effectively with a machine that responds to natural language. For a business team, mastering it makes the difference between using AI as an erratic intern and using it as a consistent collaborator. The mental model that Anthropic itself recommends is a helpful one: treat the model like a brilliant but brand-new employee who does not know your company, your clients, or your context — anything you fail to tell it, it will either make up or handle in its own way.
Has prompt engineering died in 2026?
The job title died; the competency became universal. In 2023, "prompt engineer" roles were advertised at $200,000–$300,000; since late 2024 those listings fell by around 30% and the function was absorbed into existing roles — product, operations, marketing, development. At the same time, the centre of gravity shifted: in June 2025 Shopify CEO Tobi Lütke popularised the term context engineering, and Andrej Karpathy gave it a definitive framing days later: "the delicate art and science of filling the context window with just the right information for the next step." The business lesson is concrete: the value no longer lies in clever phrasing formulas, but in what information you give the model — documents, data, examples, criteria — and in ensuring that every employee who uses AI daily shares a common baseline level of skill. With AI adoption among Spanish companies having nearly doubled in two years (from 12.4% to 21.1%, according to Spain's national statistics institute INE), that common baseline is the outstanding assignment.
The techniques that still work (backed by evidence)
The official guides from Anthropic, OpenAI, and Google converge on a surprisingly stable core:
- Clarity and specificity above all. Anthropic's "golden rule": if a colleague would be left confused reading your prompt, the model will be too. Always include the purpose, the intended audience, and the expected format.
- Examples (few-shot). The best-validated technique for controlling format and tone: Google recommends including examples "always," and Anthropic describes it as one of the most reliable ways to steer output. Two or three examples of the desired result are worth more than a paragraph of explanation.
- Structure with delimiters. Separating instructions, context, and data with XML tags or Markdown headings improves reliability — all three providers recommend it. A well-structured prompt is a document, not a WhatsApp message.
- Context first. A little-known but widely cited detail in Anthropic's documentation: with long documents, placing them at the top of the prompt and the question at the end improves performance across all models.
- Positive instructions. Saying what to do works better than listing prohibitions ("write in a formal tone" rather than "don't be informal").
- Role and system prompts, with realistic expectations. Assigning a role ("act as a chief financial officer") remains useful for tone and focus — Anthropic states that even a single sentence makes a difference — but the evidence that it improves accuracy is weak. Use it for style; do not expect miracles of precision.
The zombie techniques: what you should no longer do by default
Here is what almost no guide in English bothers to say clearly. "Think step by step" has expired as a universal habit. The Wharton report (June 2025) measured the effect of chain-of-thought on reasoning models: marginal improvements (+2.9% on o3-mini) and even negative results (−3.3% on Gemini Flash 2.5), with 20–80% more latency. A meta-analysis of more than 100 papers presented at ICLR 2025 put it in context: step-by-step reasoning helps mainly in mathematics and symbolic reasoning tasks, and contributes almost nothing to general knowledge tasks. Models in 2026 already decide for themselves when to reason; OpenAI's official recommendation is to treat them like "a senior colleague to whom you give the objective," not the procedure. Capitalised imperatives and threats are also superfluous: Anthropic warns that aggressive language such as "CRITICAL: You MUST..." triggers overreactions in current models. And those 500-word magic prompt templates circulating on LinkedIn have aged badly: the real lever has moved from phrasing to the context you supply.
The risk nobody talks about: your data inside the prompt
The prompt is also a channel for information leakage, and the 2025 data are serious: 77% of employees who use generative AI paste corporate data into the chatbot, and 82% of those interactions happen from personal accounts outside company control (LayerX, 2025). Around one in five of those pastes contains personal or financial data — meaning direct exposure to GDPR liability before the Spanish data protection authority AEPD (Agencia Española de Protección de Datos). In parallel, prompt injection is the number-one risk in the OWASP Top 10 for LLM-powered applications for the second consecutive edition: malicious instructions hidden inside documents or websites that the model processes can hijack its behaviour. The corporate response has three layers: company-managed tools (not personal free accounts), a clear policy on which data may enter a prompt, and training so that the team recognises both risks. Prompt hygiene is just as important as prompt writing.
Prompting as a legal obligation: Article 4 of the AI Act
Since 2 February 2025, Article 4 of the European AI Regulation requires every company that uses AI — whether as a provider or a user, regardless of size — to guarantee an adequate level of AI literacy among its staff: documented, role-proportionate training. And 2 August 2026 — just days from the publication of this article — marks the start of national enforcement of the regulation. There is no official curriculum: what is required is genuine training, adapted to the way each role actually uses AI, and knowing how to phrase requests, verify outputs, and avoid data leakage in a prompt is precisely the practical core of that literacy for most positions. The upside for companies in Spain: this training is eligible for subsidies through FUNDAE (Spain's state-funded vocational training scheme). Our prompt engineering training for teams is built around this dual objective — measurable productivity gains and compliance with Article 4 — and 78% of Spanish professionals already say their company should provide it.
From individual tricks to corporate asset: the prompt library
"The typical mistake is treating prompting as an individual skill — each employee hoarding their personal tricks in a notepad. Companies that get a real return treat it as an asset: a shared, version-controlled, tested library of prompts for their twenty most repeated tasks. That turns each person's luck into everyone's standard," explains Alfons Marques, founder of Technova Partners.
The implementation plan we recommend:
- Identify the 10–20 most repeated AI tasks in each department (proposals, client emails, meeting summaries, data analysis).
- Create the first version of the library: one tested prompt per task, with its few-shot examples and output format, accessible to the whole team.
- Train the team on evidence-backed techniques — and on the lesson of the "jagged frontier" from the BCG study: training must also teach when not to trust the model, because that is where the 19 points are lost.
- Review and version quarterly: models change (remember when step-by-step was indispensable?) and the library must change with them.
If you want to understand how these models work under the hood first, our guide to what an LLM is for businesses is the natural companion to this article.
Want a custom prompt library and team training that meets Article 4 of the AI Act, with FUNDAE subsidies? Tell us about your case and we will prepare a proposal.
Conclusion
Prompt engineering in 2026 is not what it was in 2023 — and that is good news. The star job title disappeared, but the competency became mandatory for the entire workforce: for productivity (the evidence of +40% quality is real when used correctly) and by law (Article 4 of the AI Act, with active national enforcement from August 2026). What works is clear and documented: clarity, examples, structure, and context placed correctly. What is surplus is equally clear: step-by-step by default, aggressive capitalisation, and magic prompts. And the real level-up does not lie in the individual prompt, but in turning prompting into a corporate asset — a shared library, a data policy, and training that also teaches people to distrust AI at the right moment.
Frequently asked questions about prompt engineering
What is prompt engineering in plain terms?
It is the practice of designing well-structured instructions for a generative AI — what you ask for, with what context, with what examples, and in what format — in order to obtain useful, consistent results. It does not require programming skills: it is structured communication with a machine, and in 2026 it is a cross-cutting office competency, not a job title.
Is it still useful to tell AI to "think step by step"?
As a default habit, no. Studies from 2025 (Wharton, ICLR) show that on current reasoning models chain-of-thought delivers marginal improvements or even worsens results, and always adds latency. It helps with mathematical or symbolic logic problems on non-reasoning models. The current manufacturer recommendation is to give the objective and context and let the model decide how to reason.
What are the risks of using prompts in a business context?
Two main ones. Data leakage: 77% of employees who use generative AI paste corporate information into chatbots, often from personal accounts — when personal data is involved, that is GDPR exposure. And prompt injection: the number-one risk in the OWASP Top 10 for LLM applications, where hidden instructions in a document or website manipulate the model's behaviour. Both are mitigated with corporate-grade tools, a data policy, and training.
Is prompt engineering training mandatory?
Specific prompting training is not named in any law, but Article 4 of the European AI Regulation has required, since February 2025, that companies guarantee AI literacy for all staff who use it, proportionate to their role and the associated risk — and knowing how to formulate requests, verify outputs, and protect data in prompts is the practical core of that literacy. National enforcement begins 2 August 2026, and in Spain this training can be subsidised through FUNDAE.
