The AI Design Thinking Coach skill for Claude

The problem this skill tackles is not that you have too few ideas. Its own opening lines already say so: the coach does not decide for itself what the user wants, and it never jumps ahead to a solution, even when the question seems to be asking for one. That is exactly where most projects fall apart. Someone has an idea, the idea becomes a plan, the plan gets built, and only at delivery does it turn out the problem was slightly different. This skill puts a gate in between. You may only move on to Ideate once there is a problem statement that passes the check, and that check is not optional.

It is written mainly for entrepreneurs, product owners, teams and consultants who are building something new and want to avoid building the wrong thing. It is part of the skill library NL that we make freely available through our AI and automation service, with no account and no sales email afterwards. If you first want to know what a skill technically is and where the file format comes from, that is covered in what are Claude skills NL.

There is one thing you should know up front, because the file itself is crystal clear about it: the skill does not replace your user research. It helps you ask questions, but it does not know your users. It writes the interview guide, it works out what to watch for if you can observe people, and it checks whether your problem statement holds up. The answers have to come from real conversations. An empty empathy map that you fill in yourself with what you think clients believe is, according to the skill, not research but an assumption with four boxes drawn around it.

Diverging and converging happen at the same time. The model deliberately alternates between widening and choosing. It says outright that most teams do both at once and end up with nothing to show for it: every idea gets judged while it is still being spoken, so the odd variants that later turn out to be the good ones never survive that first second. That is why the skill first delivers at least fifteen unfiltered ideas in Ideate, including a few that are deliberately extreme, and only after that a selection based on criteria you have named yourself.

The team is one phase earlier than it thinks. The review mode says so in as many words: given a complete file, the skill states which phase the team is really in, which is often one phase earlier than it thinks. Anyone who says they are at Prototype but cannot write a problem statement without a solution baked into it is actually at Define.

Interviews ask for opinions instead of behaviour. This is the most concrete rule in the whole file. The writing rules require that user questions be open and in the past tense: not would you use this, but tell me about the last time this happened. In review mode the skill flags every question that asks for an opinion or a prediction and replaces it with a behaviour question. The reason is simple. People are bad at predicting their own behaviour and good at recounting it.

Going backwards is treated as failure. The model is not linear. From Test you regularly go back to Define, and the file calls that a win, not a failure. In fact, in the Test phase the skill writes down in advance which outcome means you continue and which means you go back, so that decision is not up for negotiation afterwards.

This is what the difference looks like between an opinion question and a behaviour question. The lines with a plus come literally from the interview guide in the SKILL.md, the lines with a minus are written for illustration, in the form the skill specifically disapproves of.

After that comes a block of theory with the origin, and only then the method. The skill always carries out eight actions. It first asks which phase is at hand and whether earlier phases have really been completed, and anyone who wants to start at Ideate without Define gets sent back to Define first, briefly. It works through one phase at a time and delivers an artefact per phase. In Empathize it asks questions about behaviour and events, not about opinions or predictions.

In Define it formulates a problem statement following the fixed pattern of user, need, insight, and checks that no solution is hidden inside it. In Ideate it delivers breadth first and judges only afterwards. In Prototype it first decides which assumption carries the most risk and picks the cheapest way to test it. In Test it delivers a plan with criteria fixed in advance. And throughout it explicitly flags what is an assumption and what you supplied yourself.

Before anything happens, the skill runs through a mandatory input checklist with seven points: who you are designing for, what you already know about that user and where that knowledge comes from, where you currently stand in the process, what needs to be different in three months, what constraints apply in terms of budget, time, technology, regulation and existing systems, whether you have access to real users and how many, and what has already been tried before and why it did not work. It only asks about what is missing and then stops asking, so you do not get a form to fill in.

There is also a block of writing rules, and they explain why the output reads the way it reads. The file itself is written in Dutch, and so is what it produces: clear Dutch, addressing the reader with the informal you, no dashes in running text, AI in capitals, short sentences alternating with long ones, concrete over abstract, a number over an adjective. There is even a banned list of AI words and hollow superlatives. Want to set up a file like this yourself? The method is in writing a SKILL.md NL.

Finally, the file has a separate review mode for six kinds of input: a problem statement, a set of How Might We questions, an interview guide, a prototype, a test plan and a complete file. So you do not have to be starting from scratch to get something out of it. You can also throw in work that already exists and have it checked for where it rattles.

The fame comes from Brown and IDEO. Design Thinking became widely known through Stanford University's d.school and through the design agency IDEO. Tim Brown, then CEO of IDEO, described the approach in a 2008 Harvard Business Review article and in his 2009 book Change by Design. The five phases as the skill uses them are generally attributed to the d.school, which describes them in its own teaching material. In 2015 IDEO published The Field Guide to Human-Centered Design.

The roots are older than the label. Herbert Simon was already writing about design as its own way of thinking in The Sciences of the Artificial in 1969, and Nigel Cross researched how designers reason, among other places in Design Thinking in 2011. An origin like that is not a footnote. It explains why the skill treats design as a way of thinking with its own rules, not as a creative mood.

The diamonds come from the Design Council. The 2005 Double Diamond model shows when you widen and when you choose: find the right problem first, then the right solution. You can use that model when you have to explain to a client or a team why you are not choosing yet.

The five day version comes from Knapp. Sprint, from 2016, written with Zeratsky and Kowitz, compresses the phases into five days with a tight daily rhythm. This variant is deployed when there is time pressure and decision-makers at the table, with an explicit warning: do not skip the broad Empathize phase, compress it into a day with experts instead.

The How Might We question form comes from Basadur. That form is generally attributed to Min Basadur and became widely used through IDEO. In the skill it is the bridge between Define and Ideate: five to eight questions follow from the problem statement, and every question is judged as too broad or too narrow.

On top of that there are four variants you may deploy when the context calls for them. Human-Centred Design as IDEO describes it, with the three-way split of inspiration, ideation and implementation, is useful when implementation is the real sticking point. And Jobs to be Done can replace the Empathize phase when you cannot talk to users but do know their behaviour and purchase moments. There is a skill of its own for that, the Jobs to be Done Analyst.

After that, using it is simple: describe who you are designing for and where you currently stand, and the skill asks itself what is missing. If you get stuck anywhere, the full step by step plan per environment is in installing Claude skills NL, and the broader explanation of working with AI is in the knowledge base. Want your whole team to learn how to set up this kind of working method themselves? That is the subject of our AI training NL.

Double Diamond when you have to explain to a client or a team why you are widening now and only choosing later, because that model makes the two choice moments visible. The Design Sprint when there is a deadline, decision-makers are available and the problem is already reasonably sharp. Only Define and Ideate when the team already has a lot of user knowledge but keeps circling the same solution direction. And for internal processes you simply point Empathize at the employee as the user: the five phases work the same way there.

There are also situations where you are better off leaving it aside, and the file's frontmatter names two of them. The first: when the problem is technical and clear cut and the solution is already fixed. Then you do not need a design process but someone to execute, and working through five phases is wasted effort dressed up. The second is more fundamental: the skill is not a substitute for real user research. It helps you ask questions, it does not know your users. If you have no access at all to people who have the problem, this process delivers at best a better worded assumption. That is still worth something, but call it that.

One more honest limit. This coach makes your process better, not your market. If there is no problem big enough for people to pay for, that surfaces in Test, not before. That is exactly what the model is for, but it also means a well run process can sometimes end with the conclusion that you should not carry on with this. Anyone who would rather force that conclusion earlier can take the idea apart in advance with the Pre-mortem Analyst NL.

A well run design project delivers something people are actually waiting for. Then a second question begins: do those people know it exists. That is our trade. The free SEO scan shows in a few seconds where your site currently stands, and if you want to talk through what else AI can do for your business, from separate skills to full automation, we simply do that in a conversation.