Section D — Learning Path

Owners and decision-makers are called on to enable their staff to learn the new technology where it is meant to be used: in the office, at work.

This task cannot be delegated. Not to younger colleagues, who take to new digital systems quickly but do not yet have enough experience of what everyday planning demands. Nor to the BIM and IT specialists, who mastered the technology long ago. The value arises where the centre of gravity of the work lies: in the middle.

The decisive ability for using AI productively at work is not a software skill. It is older and more familiar than any AI system: the ability to formulate a task, clarify requirements and break a complex goal into workable steps. If you can delegate well, you already bring what matters to working with AI.

Nate B. Jones has defined six areas of competence that capture our own experience of recent years beautifully. Before those, three simple basics to start with:

a) Set up your workplace — try AI where you actually work, and while you are at work. Do not put it off to the evening or the weekend. Prompting has to work in the moment when your head is set to “work”.

b) Format the results — get clear about the basic rules by which the chatbot’s answer is formatted. Do not be frustrated if it does not work straight away. Practice makes perfect.

c) Prompting as dialogue — talk through the tasks on your mind with your chatbot. Discuss possible approaches. Stop the chatbot from setting to work before you have defined the objective well enough.

The six areas of competence after Nate B. Jones:

Context Assembly — the ability to assemble relevant information so that an AI system can understand and address a task precisely.

Quality Judgement — the ability to assess AI-generated results critically and recognise when they are reliable and when they are not.

Task Decomposition — the ability to divide complex tasks into manageable, clearly defined steps.

Iterative Refinement — the ability to treat a first AI result as a starting point and bring it to the quality you need through step-by-step revision.

Workflow Integration — the ability to embed AI seamlessly in everyday work rather than treating it as a separate activity.

Frontier Recognition — the ability to recognise the limits of current AI systems, and to know where they perform well and where they fail.

Our strong recommendation: these competences cannot be studied in theory. They arise only through practice — in your own working context, on real tasks, while the office is running.

The LEARNING section is meant to give you a guide into practice: the chapters that follow are structured along these six areas of competence.

Handwritten January 2026, editorially revised March 2026 – translation: Claude, August 2026.

You can memorise any number of textbooks on a foreign language. The language itself, though, is only learned by using it.

There is no instruction manual for artificial intelligence. However understandable the initial need for orientation may be, none of it counts for anything unless you find your way into practical use.

Most of the people we have spoken to and watched over recent years fail at three hurdles. AI cannot deliver, straight off, the products that everyday professional life normally expects of you and your work. The legal framework, particularly for professional use, is unclear. And the professional environment makes it hard to find a practical way in alongside the actual work.

Handwritten January 2026, editorially revised March 2026 – translation: Claude, August 2026.

With the exercises that follow, bear in mind that our perspective on AI is not necessarily yours.

This book aims to give as general an overview as possible — but it is written from the perspective of AI enthusiasts who have built up practical experience over years and now take a good deal of it for granted. The recommendations that follow are also aimed at practical use in the context of a planning practice.

Creative practitioners and designers will look for a different emphasis than AI sceptics. Newcomers will take a different route than people who have long practical experience and want to use this book to check what else there is to know.

Handwritten January 2026, editorially revised March 2026 – translation: Claude, August 2026.

Aim: set up a practice ground.

Create a working environment and make first contact with several chatbots. Set up a filing structure so you can share your results with others and assess them later yourself.

Try to work on subjects and tasks related to your daily work. But avoid projecting the pressure and expectations of your professional life into the very first tasks you take on with AI.

Learn to experiment, to play and above all to learn from failures. Anyone who observes honestly at this stage what AI can and cannot do lays the foundation for Frontier Recognition — the ability to judge the limits of a system realistically.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.1.2.1 A space to practise. AI is best learned in the environment where it is meant to help you work. For support in the office, then, practise directly in your working environment. Make appropriate time for it during the working day.

D.1.2.2 Orientation. AI is a new tool. First learn to handle small tasks with general AI applications before looking for specific software to take over larger parts of your work. Frontier Recognition begins here: if you do not know a tool, you cannot judge its limits.

D.1.2.3 Basic understanding. AI is a new language. Even though its surface is your mother tongue, you will have to learn a new way of thinking and formulating to work with these systems. Context Assembly — the ability to give a chatbot the information it needs to do the work — has to be learned. It develops through repetition.

D.1.2.4 Experiment. To understand what AI can do, changing perspective is essential. Allow yourself to work on tasks that lead to nothing usable. Iterative Refinement begins with looking at a first result — and writing the next prompt instead of giving up.

D.1.2.5 Learning from mistakes. AI will not understand what you want first time. Pay particular attention to wrong results. You usually learn more from those than from quick successes. A poor result points to an imprecise prompt — and is therefore an invitation to develop your Context Assembly further.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.1.3.1 Exercise — set up a workplace.

Make the space you need to do small exercises right in the middle of work.

Task — a place to practise: set up a separate Google email account for training purposes. Use the Google Drive access that comes with it to store training material and results. Create a thematic structure so you can follow your own progress on different tasks over time.

D.1.3.2 Exercise — steering the format of answers.

Formatting results — learn to shape the answers AI produces deliberately.

Task — understanding roles: set up an agent and give it, in the system prompt, a role and a specification for the format of its answers — a poet who only answers in verse, say, or a sullen teenager who never manages two connected sentences. Then ask: “What’s for dinner tonight?” Repeat the experiment in different configurations and compare the results.

D.1.3.3 Exercise — talking to the chatbot.

Prompting as dialogue — learn to capture and structure freely spoken content with AI.

Task — AudioChat: hold a freely spoken conversation with your chatbot in AudioChat mode, discussing what you expect from artificial intelligence. Consider which parts of your daily work AI should support and which tasks you would happily hand over to it completely. Close with the instruction: “Summarise the conversation and turn it into a structured list of topics.” Use the summary of the first conversation as an attachment for a second round.

D.1.3.4 Exercise — start with something simple.

Quality Judgement — learn to visualise a simple design goal with AI.

Task — T-shirt contest: use an image generation tool to create a white T-shirt against a neutral background. Then design the shirt with colours, patterns and appliqués until it makes a fashion statement. In a second chat, create a neutral figure and dress them in the shirt you designed. Finally, transfer the figure with their shirt into the context of a fashion shoot. Notice how your prompts have to be written for the result to match what you had in mind. Task Decomposition begins with breaking a complex image request into successive, individual instructions.

D.1.3.5 Exercise — reviewing the failures.

Frontier Recognition — learn to be clear about the circumstances in which AI does not understand what you asked for.

Task — T-shirt fails: add the failures to your presentation of the T-shirt challenge — the images that did not turn out as you wanted. For each one, note what caused it: too vague a prompt, the wrong tool, or a task the system is fundamentally unsuited to. Frontier Recognition begins with honest documentation. Arrange a session where you and your colleagues show each other your T-shirts.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

Aim: keep your own opinion.

Learn to discuss more complex content with AI and to draw a structured conclusion from the exchange at the end. What substance does the technology offer in answer to your questions? How does it respond to objections? Does the discussion turn into an exchange with a yes-man?

Use AI to form a considered opinion on various ethical questions. In dialogue with the chatbot, try to find where the line runs between a digital assistant and your own opinion. Explore ways of turning the chat history into a position paper you can use afterwards.

Frontier Recognition matters particularly here. A chatbot can supply arguments, formulate counter-arguments and name sources. Whether the arguments hold and the sources are reliable is for you to judge.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.2.2.1 Jobs. You grasp the effect of AI on your own work best by engaging with the technology actively. The most important solutions can only come from concrete knowledge of your own tasks. No outside consultant and no chatbot knows your working day as well as you do.

D.2.2.2 The profession. AI will change many professions lastingly over the coming years. Analyse how your own workload might change through AI and which processes could be automated. Task Decomposition helps here: if you can break your own work into steps, you see faster which of them can be delegated.

D.2.2.3 Energy consumption. Work responsibly and sparingly with resources as a matter of course. Be clear about where the largest consumption in your professional practice lies. It is our responsibility to use AI for socially productive solutions that justify the energy consumed.

D.2.2.4 Creativity. So far AI shows little ability to formulate an independent thought without being directed, or to set a direction of its own. Try, experimentally, to develop a feel for how much direction is needed — that is the core of Context Assembly: without enough context and a clear sense of direction, the result stays generic.

D.2.2.5 Data sovereignty. Architecture lives on precedent. How much of a design is genuinely new and how much rests on borrowed forms, typologies and references is a question everyone has to answer for themselves. Is your own creation a work to be protected from unauthorised use? Or a public contribution you are glad to see noticed and quoted?

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.2.3.1 Exercise — what work can my chatbot take off my hands?

Iterative Refinement — learn to define concrete areas of work in dialogue with the chatbot.

Task — taking stock of your work: take the result of your first discussion with the chatbot as an attachment for a new thread. Ask the chatbot to name a series of tasks it could genuinely take over from your daily work. Continue the dialogue until you have identified two or three tasks you can assess well and immediately from your own experience. Carry out the most obvious one. Could you use the result directly?

D.2.3.2 Exercise — how does AI transform my workflow?

Task Decomposition — break the previous task into smaller steps, and use them to steer the chatbot through the work.

Task — swapping roles: how would you explain the task you just worked on to an intern? Walk the chatbot through a sequence of steps as small as you can make them. Choose your words so that misunderstandings are ruled out as far as possible. What project-specific context is needed? Compare the result with your own work: what did AI do better or faster? What worse? What is missing altogether?

D.2.3.3 Exercise — how much energy do I use with AI?

Task Decomposition — learn to produce a checkable calculation with AI through a deliberate sequence of steps.

Task — energy balance: use AI to establish the standard energy consumption of your workplace in normal operation (lighting, heating, coffee machine, computer, telephony and so on). Break the result down proportionally into productive working time and standby time. Then set against it: how much text does an AI generate for the figure you arrived at (kWh)? How does the balance change with an average of one or two days working from home per week?

D.2.3.4 Exercise — one image: can AI be creative for me?

Frontier Recognition — get a clear sense of the limits of AI’s spatial imagination.

Task — design statement: give your chatbot as detailed a written description of an architectural composition as you can — ideally the opening paragraphs of the design statement from your last competition entry. Then ask the chatbot to translate the written account into one or more visualisations. How far is the result from what you intended? Try the reverse: give the AI a visualisation from one of your projects and ask for a detailed description.

D.2.3.5 Exercise — one design: can I be creative with AI?

Task Decomposition — develop a design of your own.

Task — data sovereignty: take a motif from the previous exercise as your starting point. Working in dialogue with your chatbot, develop it into as complete a design as you can: plan, elevation, materiality, colour, detail, fittings. Use only your own language. Under no circumstances quote external authors in your prompt — name no specific products, styles or existing buildings.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

Aim: generate your first content.

Learn to work out structured content that you can later use as a consolidated body of knowledge.

For use in professional life, the legal framework and duties of care must be observed without exception. Does the firm you work for already have an AI policy? If so, bring it into the exercises that follow where you can. If not, use the exercises to give yourself a framework for working with AI.

Quality Judgement is the central competence in this chapter: if you know the legal framework, you can judge which results a chatbot delivers reliably — and on which subjects an external check remains essential.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.3.2.1 Duties of care. Use AI to get clear about the most important duties of care to observe when working with it. What are you currently using AI for? What content do you upload for processing, and what do you use the generated content for? Develop a memo sheet so that your findings are understandable and available to everyone around you.

D.3.2.2 Quality assurance. Quality Judgement does not develop through a single check but through repetition. Ask the same question several times and compare the answers, and you get a feel for where a language model works reliably — and where it sounds convincing but is wrong. Responsibility for the result stays with people. No final approval without a human checking it last.

D.3.2.3 Personality rights. Pay attention to protecting individual data and rights. Do not process private information such as portrait photographs, addresses or bank details with AI unless you have explicit permission first.

D.3.2.4 The EU AI Act. Use AI to orient yourself in complex legal frameworks. In the context of your own concrete use cases, work out a summary of the most important requirements the EU AI Act places on your work. Ask your chatbot to support its answers with specific articles of the legislation — and check whether the passages it cites really say what it claims.

D.3.2.5 Copyright. When working with copyrighted content such as photographs or texts in AI systems, the author’s consent must in principle be obtained in advance. If you use AI-generated content as part of your work, do so consciously and mark where it came from.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.3.3.1 Exercise — aggregating content from a list.

Context Assembly — learn to turn a loose list of keywords into a structured paper.

Task — a checklist for data sensitivity: from memory, put together a loose list of keywords: what kinds of information do you handle in a typical project week? Texts, photographs, drawings, meeting notes, client data. Give your chatbot the definition of the four circles from Chapter C.1.16 and have it assign your keywords to them. Which of these activities are unproblematic? For which terms did copyright or personality rights come up? If you can, extend the approach to other areas of work in your practice.

D.3.3.2 Exercise — judging the quality of results.

Quality Judgement — learn to judge whether a statement generated with AI is complete and correct.

Task — fact check: choose something you are currently working on and know well — a normative requirement on your current project, say, such as the requirements for a door between a kitchen and a dining room. Ask the chatbot to produce a detailed account of all the guidelines and requirements that apply. Check what it says against your own knowledge or a source you trust. Note what was correct, what was missing and what was wrong. Then repeat the exercise with a subject you know less well.

D.3.3.3 Exercise — anonymising a photograph.

Iterative Refinement — learn to protect the personality rights of other people.

Task — retouching photographs: you have a photograph from a site meeting in which two of the client’s staff are visible in the background. Mask out the two figures in an image editor and write a prompt that restores the marked photograph with two neutral figures. Finally, ask the chatbot to turn your approach into a general procedure that could later apply across your whole practice.

D.3.3.4 Exercise — questioning a full text.

Context Assembly — learn to take in complex professional content indirectly, through AI.

Task — three questions for the EU AI Act: collect every question you or your colleagues have about the EU AI Act. Upload the full text of the Act to your chatbot and ask: “Answer each of my questions and support your answer with the specific article you are relying on.” Then ask it to name, building on your list of questions, further questions and requirements that the EU AI Act creates for a planning practice. From that, extract a memo sheet of mandatory and optional requirements.

D.3.3.5 Exercise — quoting an existing work in a new project.

Iterative Refinement — learn to carry the design qualities of your own work over to a new project.

Task — transferring a design: find a photograph of one of your genuinely successful buildings. Take the image apart into its most important design components: the composition of colour, the materiality, the key formal details or fittings. Find a suitable new project the design could carry over to well. Try to guide the AI in projecting the design onto the new object or into the new space. Review the surprise successes and the failures as a group: who owns the result?

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

Aim: build AI into your daily workflow.

“What requirements apply to the external works? Check the fire strategy. And compare its provisions with the overarching building regulations.”

Do not put this question off. Give the task to AI in the moment when you have worked out in your head how it needs to be solved.

Our experience is that AI works best in professional life when it is used in the moment. Just as you jot an urgent task on a notepad, productive prompting works best exactly when your brain has given the task its first concrete shape.

Resist the impulse to make a note reminding yourself to “discuss that with the AI the day after tomorrow”, now that you have just pictured the task vividly. Experience shows the thought has evaporated once you have put the task off into the future.

Workflow Integration is the competence of this chapter — and it is less a question of technology than of habit.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.4.2.1 Context. AI needs a link to your professional context. Learn to formulate small tasks that sit as close as possible to the ordinary demands of your working day. The more specific the context you give the chatbot, the more usable the result. The chatbot knows nothing about your practice, your projects or the basis on which you decide things — until you tell it.

D.4.2.2 Currency. No AI application and no new digital planning tool is a finished solution. Development continues, and new technical possibilities appear every month. Frontier Recognition also means this: limits that held six months ago may have been overcome today — and the other way round.

D.4.2.3 Analysing the process. AI image generators respond to language — but not always as you expect. The first result is a starting point, not a finished product. Iterative Refinement means not breaking off the dialogue when the first image is wrong, but reading the result — and writing the next prompt more precisely.

D.4.2.4 Building your own. One central benefit of AI lies in developing specific applications that are available to you and your colleagues again and again in daily work. The first system prompt is rarely the best. It gets better through repeated use and step-by-step revision.

D.4.2.5 Assessing the benefit. Analyse together the workflows where you have identified potential for AI to improve things. A sober balance — where AI actually saved time, where it created more work — is the basis for everything you plan next.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.4.3.1 Exercise — turning an email into a work plan.

Workflow Integration — learn to bring AI directly into your running project correspondence.

Task — project correspondence: Step 1: from an email of the last few days, copy the passage that sets out one or more tasks currently due. First ask your chatbot to turn the passage into a list of keywords for the tasks. Step 2: ask the chatbot to work out what exactly has to be done for each point. Step 3: give the chatbot a description of the people involved and let it decide who exactly has to do what. Then check how precisely the chatbot was able to interpret and assign the tasks.

D.4.3.2 Exercise — design an accessible WC.

Frontier Recognition — learn to recognise the limits of AI on a concrete design task.

Task — accessibility: Step 1: have a dimensioned floor plan of an accessible WC produced to current German building law. Step 2: have a list produced of the most important dimensional rules for an accessible WC under current German building law. Step 3: compare the written requirements with the graphic result. Can you find a way to get the AI to produce a dimensionally correct drawing? Try the reverse: can you simplify the task far enough that the AI does produce a dimensionally correct plan?

D.4.3.3 Exercise — design an interior step by step.

Iterative Refinement — learn to express your design intent through AI image generators.

Task — my favourite room: Step 1: take a photograph of any interior. Ask the AI to remove all furniture and decoration in the first step. Step 2: furnish the room step by step — from the floor material to the wall colour, to the arrangement of the main pieces of furniture, to the final addition of pictures and other details. Step 3: from the best result, derive further illustrations and variants deliberately. How many steps did it take before the result matched what you had in mind? Note at which point you stopped correcting the AI — and whether that was because the result was good, or because you had adjusted your expectations.

D.4.3.4 Exercise — build a checklist for the next submission.

Task Decomposition — learn to support yourself with small applications in daily work.

Task — planning application checklist: Step 1: discuss your current project with your chatbot. What deliverables do you owe in the current service phase? Step 2: discuss which essential sub-steps and preconditions are required. Have the result turned into a structured task list in Excel. Step 3: with the chatbot, convert the Excel table into an interactive HTML page where you can set each step to “in progress” or “done”. An Excel table cannot be opened in a browser — an HTML page can, on any device, without extra software.

D.4.3.5 Exercise — have a bill of quantities produced.

Quality Judgement — learn to test the quality of an AI-generated technical text against a reliable original.

Task — image to text: Step 1: from one of your recently completed projects, find the bill of quantities for a one-off component — a fitted kitchen, an elaborately detailed washbasin unit or a bespoke spiral staircase, say. Step 2: give your chatbot a photograph of that component and ask for a detailed description with all materials and points of detail. Step 3: ask the chatbot to turn the detailed description into a bill of quantities. Finally, compare the AI’s version with your original.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

Aim: bring some order to the AI sprawl in the practice.

Without an AI guideline and without talking to staff directly, sprawl sets in quickly, in the office and on the project. As browser-based software, countless offerings — often free, which means you pay with your data — are available through any firewall on every computer in the practice. What sensitive content colleagues are processing with systems without a business licence, outside the European legal area, cannot be overseen.

Change management is not desirable in this context but essential.

This chapter is addressed to those who carry responsibility in a practice. Workflow Integration and Context Assembly are not individual competences here but organisational ones: how do you make sure the whole team uses AI sensibly, lawfully and productively? Frontier Recognition applies to leadership teams too — you have to know which processes suit AI and which do not before you take decisions.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.5.2.1 A leadership team. The use of AI has to be matched against the aims of the business. An AI leadership group is indispensable for that: one that accompanies the methodical integration continuously and secures the organisational framework it needs. Context Assembly applies at leadership level too: if you cannot state the aims and constraints of your practice clearly, you will not use an AI system sensibly either.

D.5.2.2 Willingness. Identify the people in your organisation who show interest and willingness towards the new technology. Encourage curiosity and a readiness to learn across the team. Share successes and experiences in the group. Iterative Refinement applies to teams as well: the first attempt will not be perfect — and that is fine.

D.5.2.3 Budgeting. Record the tools you want to test. On that basis, work out an annual budget for the licences you want to buy for trials. From the initial phase, derive an estimate of how much time systematic testing needs.

D.5.2.4 Communication. Make sure the introduction of AI is communicated transparently across the whole staff. Record which AI software is in use and write internal rules for using these new tools. Without clear communication, uncertainty grows — and uncertainty encourages exactly the sprawl you want to avoid.

D.5.2.5 Risk analysis. Analyse together the various software solutions you intend to bring in. Assess the risk if a piece of software should stop being available. Are there alternative providers? In what data formats are the results stored? Frontier Recognition here means: not every new solution is as stable and permanently available as its provider promises.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.

D.5.3.1 Exercise — turning a transcript into minutes.

Frontier Recognition — learn to produce minutes from a transcript.

Task — the inaugural meeting: use the material developed in D.2 and D.3 — duties of care, quality assurance, data protection, copyright, the EU AI Act and ethical guidelines — as the basis for the agenda of the first meeting of your AI leadership team. Make sure the conversation is transcribed. Then use the transcript to have minutes produced — with results, open questions and agreed next steps. Watch which configurations and which content work, and under what circumstances the spoken word carries so many implications and outside references that the AI can no longer follow the substance.

D.5.3.2 Exercise — making a second “product” from one piece of content.

Workflow Integration — identify areas where AI can be integrated easily and immediately.

Task — launching AI in the practice: use the agenda, the minutes and the draft AI policy from the previous exercise as your source material. Ask your chatbot: “Turn this into a short piece for our intranet or an email to the staff — factual, welcoming, without jargon. What did the leadership team discuss? What is changing? Where can staff get involved?” Then check what effect it has among your staff.

D.5.3.3 Exercise — vibe coding: building small applications into daily work.

Workflow Integration — learn to write small applications with AI that you can use again and again in daily work.

Task — what does it cost: in dialogue with your chatbot, produce an overview of the AI tools you have used in recent weeks. For each one, add: monthly cost, estimated time saved per week, number of staff for whom a licence would make sense. Then ask: “Calculate a simple annual budget from this and draft a first proposal for my budget planning.” From that calculation, have a small variable application built with which you can test different assumptions.

D.5.3.4 Exercise — building an agent for internal communication.

Context Assembly — learn to build a simple knowledge instance.

Task — the knowledge instance: gather the guidance produced in the “Obligations” exercises together with the minutes of the inaugural meeting of your AI leadership group. Ask your chatbot: “Help me develop a system prompt for an AI instance that informs our staff about using AI in the practice.” Upload the collected documents as its knowledge base and set the instance up as a Custom GPT, Gem or Agent. Test the agent as a group with concrete questions.

D.5.3.5 Exercise — building an agent for legal assessment.

Context Assembly — learn to build a simple knowledge instance.

Task — software on the test bench: work out the criteria a software provider has to meet to count as GDPR-compliant. Where is data processed? Under which legal framework? Is there a GDPR-compliant usage agreement? In a second step, use that as the basis for the system prompt of an agent that checks software providers against these criteria automatically. Test the agent as a group with the software your practice normally uses.

Conceived October 2025, revised by hand March 2026 – translation: Claude, August 2026.