Every expert I talk to about building a skill asks the same three questions, usually in the same order. How long does this take me? What do I actually have to hand over? And — said more politely, but this is the real one — how do I know you're not just going to take my life's work and walk off with it?
Fair questions, all three. The biggest barrier to a good expert building a skill isn't reluctance. It's that nobody's shown them what the process actually is, so it sits in their head as a vague, slightly threatening unknown. So let me pull the curtain all the way back and show you exactly how one gets built, using a methodology we've put through the whole machine: Ryan Singer's Shape Up.
No mystery, no hand-waving. Here's the pipeline, stage by stage, what happens at each one, and what it produces.
The shape of the thing
A skill goes through six stages. Each one has a job, and — because this matters more than people expect — each one runs on the AI model best suited to that job. Research wants a different kind of model than extraction; extraction wants a different one than copywriting. Matching the model to the work is half of why the output is any good.
The stages are: evaluate, craft, capability, build, test, market. Candidate goes in one end. A tested, live skill comes out the other. Walk through them with me.
1. Evaluate — is this methodology even skill-able?
Not everything should be a skill, and saying so up front saves everyone a lot of wasted effort.
The first stage researches the methodology and asks a blunt question: is there real, teachable judgment here, or is this just vibes? A genuine methodology has structure you can extract — thresholds, a sequence, decision points, a way of telling good from bad. A motivational philosophy usually doesn't. This stage runs on a model tuned for research synthesis, because the job is breadth: read everything around the methodology, map what's there, and grade it honestly.
For Shape Up, the answer came back strong. There's a real system in there — appetites, shaping, the binary fit check, the way Singer thinks about risk and unknowns. It's specific, it's teachable, and crucially it produces different advice than generic product management. That last part is the whole test. If an extracted skill says the same thing base AI already says, there's no point building it.
What you provide at this stage: essentially nothing yet. We do the research. You get a brief back telling you what we found and whether it's worth proceeding.
2. Craft — extract the actual judgment
This is the heart of it, and it runs on the most capable model we use, because extraction is the hardest cognitive job in the pipeline.
Crafting is where your methodology gets turned from prose into decision logic. The thresholds you carry in your head — what "good enough" actually means — get written down explicitly. The questions you ask get sequenced into the order you actually ask them, because the order is part of the expertise. The patterns you recognise get named. The specific format you deliver verdicts in gets captured, because the way an expert presents a result is itself part of the method, not an afterthought.
For Shape Up, this is where the skill learned things like: when every proposed shape fails on the same requirement, the problem isn't the shapes, it's the requirements — go back upstream. That's not in any one chapter as a tidy rule. It's the kind of judgment that's distributed across the whole methodology and the practitioner's instinct. Pulling it out and making it explicit is exactly what this stage exists to do.
What you provide: your source material — the book, the talks, the articles, whatever captures your thinking. The richer the source, the better the extraction.
3. Capability — design how it actually engages a person
A pile of extracted judgment isn't yet a good experience. The capability stage designs the engagement — how the skill meets a real person with a real, half-explained problem.
This is where a lot of AI products quietly fail, and it's worth being honest about why. People don't show up with clean inputs. They show up confused, defensive, mid-mess. A skill that only works when you speak to it in the methodology's own vocabulary is a skill that works in a demo and breaks in real life. So this stage designs for the messy middle: how the skill opens, how it draws out the information it needs, how it handles someone pushing back, how it produces something you can actually use at the end rather than just a nice chat.
This stage also runs on the top-tier model, because designing a genuine practitioner engagement is judgment-heavy work in its own right.
What you provide: ideally, some back-and-forth on how you handle people in your domain — where they get stuck, how you unstick them. That tacit stuff is gold, and we'll come back to how we capture more of it.
4. Build — assemble it
The build stage is the most mechanical, and it runs on a faster, more efficient model because the hard thinking has already been done. This is assembly: taking the extracted methodology and the engagement design and putting them together into the actual skill — the thing that runs.
There's not much drama here, which is the point. By the time you reach build, the decisions that matter have been made. This stage just constructs the artifact cleanly and consistently.
What you provide: nothing. This one's on us.
5. Test — prove it actually works
I care about this stage more than any other, because it's the difference between a skill and a vibe with a price tag.
Testing is where we find out whether the skill actually carries the expert's judgment or just gestures at it. We run methodology fidelity testing — does it follow the real method, or does it drift? We grade multi-turn conversations — does it stay coherent across a long session, or fall apart? We run full journey tests of 30 to 50 turns — because the real test of a skill isn't a one-shot answer, it's whether it holds the thread through an entire piece of work. And we run it across a range of personas, because a skill that only works for one type of user isn't done.
The Shape Up numbers tell the story. Against base Claude on real scenarios, the skill scored 18% higher overall — but the lift concentrated where it counts: voice, mastery, methodology, the expert dimensions. And the multi-turn result is the one I always point to. In a ten-stage shaping session, base Claude lost the methodology thread after about turn five and reverted to generic product advice — it completed maybe 60% of the journey holding the method. The skill held it through all ten stages. That gap — between something that's good for one reply and something that stays expert across a whole engagement — is the entire reason testing exists.
This stage runs on a top-tier model too, because judging quality is as hard as producing it.
What you provide: nothing required, though if you want to throw your own tricky scenarios at it, this is a great moment.
6. Market — make it findable
The last stage writes the positioning — how the skill is described, who it's for, the copy on its page. It runs on a lightweight, fast model, because good marketing copy is a craft that doesn't need the heaviest reasoning, just clarity and discipline.
What you provide: a review, if you want one. It's your name on it, so you get the final say on how it's presented.
"But the pipeline only gets you to 70%"
Here's the part I won't pretend about, because the honesty is the point.
The pipeline I just described is good. It gets a skill to roughly 70% of what you, the actual expert, would call a faithful representation of your judgment. That's a genuinely useful skill. But the last 30% — the real tacit stuff, the calibration you can't quite articulate until you see the AI get it slightly wrong — that doesn't come out of an automated pipeline. It comes out of you.
So for experts who want to go all the way, there's a second mode. The pipeline runs first and does the heavy lifting. Then a guided refinement process — we call it working with the Director — shows you what the pipeline captured and walks you through the gaps. It opens with something like: "I found 14 frameworks and 8 decision patterns in your work. There are 3 areas where your real-world practice clearly goes deeper than the book — let's get those." And then you refine, targeted, on exactly the places where your hands know more than your book ever said.
Production mode gets you to 70% on its own. Seventy percent is a real product. But only you, in the loop, gets it to 100% — to the point where you read the output and think, yes, that's actually what I'd say. I'd rather tell you that plainly than oversell the automation.
The two questions I haven't answered yet
I said experts ask three things. I've shown you the how long and the what do I provide — most stages ask nothing of you, a couple ask for your source and a bit of your time. The third question, the real one about IP, deserves a straight answer.
Your methodology never leaves the server. For any skill carrying valuable IP, the premium delivery keeps the actual method server-side — the user receives the output of your judgment, never the judgment itself in a form they could copy. Your expertise leaves the building as advice, not as the source. That's protection by architecture, not by a hopeful clause in a contract.
And what you earn: 70% of the revenue, every time your skill is used. Not an advance you burn through once. A share that pays out with usage, across every situation your expertise touches, while you're doing something else entirely. The methodology is the whole product — so the lion's share goes to the person who built it. Everything else is infrastructure.
Where this actually leads
Here's the thing I most want a first-time creator to understand: this first skill is not the destination. It's the on-ramp.
A skill captures one methodology — one book, one framework. It's excellent at that, and for many experts one great skill is plenty. But if you've got a real body of work — several books, years of talks, a whole evolved point of view — a single skill is a slice of you. The real destination is the agent: your entire body of work, unified, interconnected, answering not "apply this one framework" but "what would you think about my situation?"
You don't build that on day one. You build your first skill. Then a second. And at two or more, the platform can unify them — extract the beliefs that run through all your work, wire the frameworks together, synthesise one voice across the lot, and stand up a living agent that grows as you feed it new thinking. The skill is the textbook. The agent is you, on call.
But that's a later article, and a later decision. For now the honest pitch is simpler than the fear that's been sitting in your head. Building a skill is a defined, six-stage process. Most of it we do. The part that needs you is the part only you can give — your judgment, and, if you want the last 30%, an hour or two in the loop. Your IP stays protected. You earn the majority share. And the thing you end up with isn't a summary of your work gathering dust. It's your expertise, applied to real problems, earning for you, at a reach you could never hit one hour at a time.
That's worth a couple of hours of your judgment. I'd argue it's the best couple of hours an expert can spend right now.
Cheers, Adam