Practice LibraryEngineering and technology2026 edition
One Hundred Prompts for Engineering
From specification to CE file, demonstrably correct.
After payment you receive a download link for the PDF by email.
Why now
There is not a single standard number in the hundred prompts in this book. No article number, no limit value, no safety distance, no Performance Level, not even as an example. That is not an omission but the core of the design: a prompt is a template that gets copied, and an example inside a template is sooner or later taken over verbatim into a file that carries your signature. We ask a language model to fill in nothing it cannot account for. We asked the same of ourselves.
From the book
Read real pages before you decide: the method, complete prompts and the safety framework.
A selection from the book. Click to enlarge.
Free sample
Read 25 real pages first. Then decide.
We are not asking you to take our word for it. Take the sample with you and judge for yourself what you are buying: how the book is built, one complete prompt with all its fields, the safety framework in outline, and the source referencing we apply on every page. Not a summary and not a brochure, but real pages from the edition.
Why give this away? Because a prompt you do not trust is worth nothing to you. If you recognise your own way of working here, you buy the rest with confidence. If you do not, it has cost you nothing.
No email address, no account, opens straight away. The last page holds a link back to this page, so you can always find your way here again.
Table of contents
- 9.1 Specifications and requirements, prompts 1 to 14
- 9.2 Risk assessments, prompts 15 to 28
- 9.3 Technical reports and calculation records, prompts 29 to 40
- 9.4 Conformity, CE and the declaration of conformity, prompts 41 to 54
- 9.5 Instructions and user information, prompts 55 to 66
- 9.6 Test, validation and acceptance plans, prompts 67 to 79
- 9.7 Verification and source binding, prompts 80 to 90
- 9.8 IP protection and safe data input, prompts 91 to 100
Around these: the problem and its urgency, the context in law and science, the BNK KAM position, the decision frameworks, practice and case studies, checklists and worksheets, frequently asked questions, pitfalls and misconceptions, a glossary and the full list of sources.
What this is for
NEN-EN-ISO 13849-1:2016, the standard for the Performance Level of safety-related parts of control systems, is listed in the catalogue of NEN (the Dutch standards body) as withdrawn on 1 June 2023, replaced by the 2023 edition. That edition went from 86 to 152 pages and costs EUR 281.00 excluding VAT. Ask a language model for a safety justification and you may get back a reference that looks exactly like a good reference. The form is right. Whether the content is right is not stated.
Standards from EN, ISO, IEC and NEN are protected by copyright and sold per document. There has never been a freely searchable corpus of standard texts from which a language model could have learned their content. That does not mean standards are out of reach: in Case C-588/21 P the Court of Justice of the European Union (Grand Chamber, 5 March 2024) held that a harmonised standard whose reference has been published in the Official Journal forms part of EU law, and annulled the Commission's refusal on the ground of an overriding public interest in disclosure. Accessible is not the same as searchable. Everything a model says about the content of a standard is reconstructed from secondary material, and that reconstruction is formally indistinguishable from a genuine reference.
Anyone who adopts a sentence produced by a language model unchanged signs for the content of that sentence. The model signs for nothing. What sets this apart from other errors is the moment of discovery. An invented calculation fails at the first check sum. An invented reference to a standard fails only when a notified body or a market surveillance authority asks for it, and by then the signature is already on the document. This book moves that discovery to the cheapest moment, the moment you are still drafting.
What's inside
- One hundred fill-in prompts in eight categories: specifications and requirements (14), risk assessments (14), technical reports and calculation records (12), conformity, CE and the declaration of conformity (14), instructions and user information (12), test, validation and acceptance plans (13), verification and source binding (11), IP protection and safe data input (10).
- A fixed safeguard paragraph in every prompt: it forbids the model to fill in a standard number, article number, limit value, tolerance, safety distance, Performance Level or SIL level of its own accord, and has everything that does come from your own material marked with [TO BE VERIFIED]. Written as text rather than as an API call, so it works with any model.
- Five fixed fields per prompt: Purpose, When to use it, The prompt, How to fill it in, and a Safety note. For risk assessments, conformity and test plans that safety note is a prohibition rather than advice, because a responsible person signs there and that signature cannot lawfully be delegated to software.
- The regulatory picture at the cut-off date: the Machinery Regulation (EU) 2023/1230, applying from 20 January 2027 and repealing Directive 2006/42/EC on that date; the revised product liability Directive (EU) 2024/2853, to be transposed by 9 December 2026 and expressly treating software as a product; and the Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026.
- Forty-eight sources, each actually retrieved and read, then independently re-checked for existence, title, publisher, year and whether the source really covers the claim.
Who it's for
For engineering, technical and QHSE managers, senior design engineers, systems and requirements engineers, and technical authors. For the person who signs and is answerable afterwards. The book does not explain your profession, it explains where the output of a language model runs aground on it.
Not for anyone looking for an AI system that takes over the assessment. That system does not exist, and this book does not pretend otherwise.
The temptation is to think that an invented reference to a standard is a rare accident. The research says otherwise. Of the citations produced by a language model, 55 per cent (GPT-3.5) and 18 per cent (GPT-4) respectively were wholly fabricated, and of the references that did exist, 43 and 24 per cent respectively contained citation errors (2023). In one medical study only 7 per cent of references were both real and correct (2023). On specific, verifiable legal questions, models hallucinated in 58 to 88 per cent of cases (Journal of Legal Analysis, 2024). And across 576,000 code fragments, models referred to 205,474 unique package names that do not exist (2025).
Note what those figures have in common. In every case the source is either public or trivial to check: a DOI, a ruling, a package name in a registry. Standard texts are neither. Checking them costs time, money and access. That is precisely why this is the domain where checking happens least often, and therefore the domain where the error survives longest.
Why BNK KAM
Because we held ourselves to our own rule. We write that a model must not fill in standard numbers because it has not read the standard texts, and for exactly that reason there is not a single number in our hundred prompts. We have not read those texts either. Standard numbers do appear in the framework chapters, but only as public catalogue data with a source reference, a withdrawal date, a page count, a price, never as a rendering of the content of a standard.
The same precision applies to how we describe the problem. Nowhere do we write that standards are secret or inaccessible, because that is untrue. What never existed is a searchable corpus.
And it applies to the 48 sources. Each one was retrieved, read and tested again in a separate round, including on the question of whether the source genuinely covers the claim. That is precisely the act this book teaches you to build in.
Proof and sources
2026 edition, research completed on 8 August 2026. Approximately 108,700 words, 100 fill-in prompts, 48 sources with full references, 5 original diagrams and original photography. Digital publication, available immediately after purchase.
Frequently asked questions
Is this not simply "check everything AI says"?
That is the conclusion, not the method. The method is that a prompt is built so that the model does not fill in the risky fields at all, and that everything coming from your material comes back marked with [TO BE VERIFIED]. You then check a defined list instead of running text in which you cannot see what was invented.
Why should I not ask the model for a standard number?
Because you get an answer that looks correct but is not based on the standard text. A model relying on older secondary material can point you with full confidence to a withdrawn edition. IEC 62061:2005 was withdrawn on 22 March 2021, NEN-EN-ISO 13849-1:2016 on 1 June 2023.
Does this book take work away or add work?
It moves work. You spend less time structuring, phrasing and keeping things consistent, and more time verifying a limited number of marked points. The difference is that afterwards you can show where that verification took place.
We work with confidential design data. Can we use this?
That is what the eighth category is for: ten prompts and the accompanying framework chapter on IP protection and safe data input, on what you do and do not enter and how to abstract without making the task useless.
Does this book replace buying the standards themselves?
No. You buy the standard texts from NEN or the relevant standards body. This book makes sure the output refers to what you supplied yourself.
Which jurisdiction does this cover?
The framework is European: the Machinery Regulation, the product liability Directive and the AI Act apply throughout the EU. Where a Dutch body appears, such as NEN (the Dutch standards body) or the Nederlandse Arbeidsinspectie (the Netherlands Labour Authority), that is stated explicitly. The book is available in Dutch and English.
This product does not contain a single invented source, figure or citation. What we teach you to demand of the machine, we first demanded of ourselves.
