The AI Business Model Canvas Coach skill for Claude

You download the zip at the top of this page, drop it into your Claude environment, and from that moment Claude works the canvas out with you using this logic. What a SKILL.md actually is and why the format stays this simple is explained in what are Claude skills NL.

The problem the skill solves rarely sits in the canvas itself. Anyone can draw nine boxes. The problem is that an empty box asks you nothing. You write "SMEs" under customer segments, "good service" under value proposition and "social media" under channels, and now you have a canvas that looks complete and says nothing. Those are not answers, they are categories, and nobody challenges you on them. The skill does. On "SMEs" it keeps asking until you have a group for which you demonstrably do something different. On a value proposition it asks for the words the customer actually used. And for every box you fill in, it wants to know whether you know it or assume it.

Founders, freelancers and teams are the readers it is written for: people setting up something new, or people who want to understand where the money in an existing business really comes from. It is part of the skill library NL we make available for free from the AI and automation service, with no account and no sales email afterwards.

One thing you should know beforehand: the canvas orders your assumptions, it does not prove them. Osterwalder and Pigneur say so themselves. A canvas that looks beautiful and rests entirely on guesses is a beautifully ordered guess. That is why the output of this skill does not end with the canvas but with the three riskiest assumptions and a test you can run this week. Usually that test is a conversation, not a research report.

The boxes are filled with categories. "SMEs", "founders", "good service", "quality", "word of mouth": these are not answers, they are drawers where answers belong. A customer segment is only a segment once you do something different for it: a different message, a different channel, a different price or a different delivery. If nothing differs, it is not a separate segment but the same group with another label. On every segment the skill explicitly asks what is different, and it only moves on once there is an answer to that.

The canvas was filled in from left to right. The way you read it, in other words. Starting with key partners means choosing partners before you know what you have to deliver. Everything left of the value proposition is a consequence: your resources, your activities and your partners follow from what you promise, and what you promise follows from who you promise it to. That is why the skill reverses the order and starts on the right, at the customer.

Assumptions are written down as facts. This is the most dangerous one, because it is invisible. In a filled-in canvas, a line you measured looks exactly the same as a line you made up. Half a year later nobody remembers which was which, and policy gets built on a guess from one afternoon. Hence the labels [KNOWN] and [ASSUMPTION] behind every line, and the rule that [KNOWN] is only allowed if you can name a source: a conversation, an invoice, a measurement, an existing client.

The nine blocks were never compared with each other. Separate filled-in boxes say nothing. The model is a system: what you organise on the left causes cost, what you deliver on the right brings in money, and those two are supposed to stand in proportion to each other. In shape and in rhythm too, because fixed monthly costs against unpredictable one-off assignments is a business model with a hole in it. The fit check in this skill exists for exactly that.

The empty blocks were quietly filled in. A canvas is supposed to have blocks you do not know yet. That emptiness is itself useful information. But both people and AI tend to plug the gap with something plausible, and that removes exactly the signal that mattered. So the skill does not fill in an empty block, it names it.

In diff form, with lines from the skill's own example: what goes out and what comes back in its place.

Then comes the core: nine steps in a fixed order, and the rule that it only moves on once the previous step has an answer that is concrete enough. Step 1 is customer segments, and it says explicitly that an answer of "everyone" or "SMEs" does not get written down but gets questioned further. Step 2 is the value proposition, where the skill asks for the pain the customer names in their own words rather than your product's features.

Step 3 is the channels, and those are asked about in four phases: awareness, purchase, delivery, aftercare. That is deliberate, because most models break down at the phase nobody had filled in. It carries on like that up to step 9, the cost structure, where costs get tied to the key activities and fixed and variable are split apart.

After those nine steps comes the fit check, and that is the real work this skill does. Five checkpoints: does every channel connect to a segment that is actually found there, does every segment have a revenue stream opposite it and every revenue stream a value proposition opposite it, do the key activities fit within the time and people available, does the cost structure stand in proportion to the revenue streams in shape and in rhythm, and is there a block that stayed empty or was only filled with a generality.

Before anything is filled in, the skill runs through a mandatory input checklist: what you do in two or three sentences, whether the business already exists, who has paid so far and for exactly what, how many people are involved and how many hours are available, your prices, your biggest cost items, what you are making the canvas for, and whether any customer conversations have already taken place. It only asks about what is missing, and at most three questions at a time, so you get a conversation and not a form.

The output format is fixed in six parts: the heading, the nine blocks with a label on every line plus a line for "still open", the canvas at a glance on half an A4, the fit check, the three riskiest assumptions each with a test for this week, and finally a block "what is missing" with the empty blocks and the open questions. That last block is never left out. The file also has a separate review mode for an existing canvas, with six fixed checks including the segment check and the missing connections. And it closes with a source list of six titles, from Osterwalder to Fitzpatrick.

The canvas comes from Alexander Osterwalder. In 2004 he earned his doctorate at the University of Lausanne with a thesis describing an ontology of business models: a structured way of naming what a business model is made of. Together with Yves Pigneur and a large group of contributors, that became the 2010 book Business Model Generation. It carries the definition this whole skill rests on: a business model describes how an organisation creates, delivers and captures value. Three verbs, and each of the nine blocks belongs to one of the three.

The value proposition later got its own book. Value Proposition Design from 2014, by the same authors with Greg Bernarda and Alan Smith, works the middle block out into a separate canvas with customer jobs, pains and gains on one side and your offer on the other. This skill deliberately keeps that separate: it asks for the promise in the customer's own words, and for picking that promise apart there is a dedicated Value Proposition Canvas skill NL.

The split between knowing and assuming comes from the lean line. Steve Blank described in The Four Steps to the Epiphany from 2005 why a startup is mostly a collection of unproven assumptions, and that you test those assumptions outside the building, not in a meeting room.

Eric Ries turned that into a working method in The Lean Startup from 2011: the smallest possible test that can knock down an assumption. That is exactly why the output of this skill ends with three assumptions and a test, not with a complete canvas as the endpoint. Anyone who wants to use the model in its lean form, with problem, solution and unfair advantage instead of partners and resources, can turn to the Lean Canvas Strategist skill.

The test is almost always a conversation. Rob Fitzpatrick describes in The Mom Test from 2013 how to run a customer conversation without steering the answer: ask about what someone has done, not what they would do. That rule is built into the skill, because testing an assumption on someone who likes you only gets you confirmation. For that conversation itself there is the Mom Test Coach skill.

Even without installing the skill you can improve your own canvas with these principles: start on the right at the customer, label every line as known or assumed, and at the end lay the nine blocks side by side instead of looking at them one by one.

After that, using it is simple: describe what you do in two or three sentences, and ask for your canvas. The more concrete you are about what has already been paid and by whom, the less stays marked [ASSUMPTION]. Want to learn to write this kind of instruction file yourself, for example for a model we have not built, then writing a SKILL.md NL explains how the frontmatter and the structure work. And the broader explanation of working with AI is in the knowledge base.

If the question is really about your target group, only the right-hand side is worked through, and pushing on to key partners is a waste of time. If you have several services, it builds one canvas per revenue model rather than one canvas for the whole business, because two revenue models in one canvas always produces an unreadable middle. And if the canvas is going to a third party, a bank for example, every assumption marker stays in.

There are also situations where you are better off leaving it be. The main one: the canvas is no substitute for customer research. Anyone who has never spoken to someone from their target group gets back a neatly ordered collection of guesses, and the skill says so itself. Talk first, then order. Financial forecasts, investment memoranda and grant applications with formal requirements are a second boundary: there the application format decides the structure, and you need an accountant, not a coach. And a third: if the real problem is that you do not have enough enquiries, a better canvas does not change that. Then the work that comes first is getting found, not ordering.

One more honest boundary: the skill makes your model legible, not viable. If the price is too low for the costs underneath it, the fit check shows that, but it does not fix it. What belongs in the place of [FILL IN] next to your prices is the domain of the Pricing Strategy Coach skill NL, and whether your offer actually matches what people want is the question of the Jobs to be Done Analyst skill. The two are deliberately kept separate.

A canvas that adds up shows where your money comes from, but there have to be people who can find you. The fit check surfaces the same point in almost every first canvas: the channel is missing. The free SEO scan shows in a few seconds where your site stands today, and for many businesses that is the channel closest at hand. In the guide which strategy model when NL you can see how this model relates to the others. And if you want to talk further about what else AI can do for your business, from loose skills to full automation, we simply do that in a conversation.