Every good loaf has a recipe. So does directing AI.
Essays with sources, techniques worth stealing, and what shipped this week — everything we
learn about directing AI, written down. No hype, numbers named, recipe shown.
the blog · fresh bakes
Read the recipes.
the evidence3 Sep 20264 min
A billion users, a 62% premium, and the 4% who train
The AI adoption curve is the steepest in consumer-technology history. The skill curve never left the ground. The gap between those two lines is the clearest market case a maker will ever get — here it is, number by number.
By mid-2026, around 1.1 billion people used ChatGPT every month (Statista), and Microsoft estimates roughly one in six humans now uses generative AI. It took Instagram two and a half years to reach 100 million users; ChatGPT did it in two months. Among students, use is close to universal — 86% report using AI in their work.
Now the other line. Only 1% of organisations describe their AI use as mature (McKinsey) — while 92% plan to spend more. Surveys through 2025 found 54% of workers call AI skills critical to their career, and about 4% are actually training them. A fifty-point gap between knowing it matters and doing something about it.
The market has already put a price on that gap. PwC's AI Jobs Barometer found roles requiring AI skills advertised a 56% wage premium in 2025 — 62% in 2026. Lightcast, across 1.3 billion job postings, measured 28%, about $18,000 a year; postings asking for two or more AI skills carried 43%. In the UK, the premium for AI skills now exceeds the advertised premium for a master's degree. And productivity growth in the industries most exposed to AI is running about 4× the least exposed (PwC).
One honest caveat: those premiums are measured on what employers advertise, not on individual raises — and at the company level most generative-AI pilots still show no measurable profit impact. Which is the point: the tools are everywhere and the direction is missing. The value is going to the people who can prove they supply it.
A billion people use AI. A rounding error can direct it well. And until now there was no way to tell them apart.
That last part is what Makery is for — not a course about AI, but a verifiable record of how you direct it. The premium belongs to the market; the proof can belong to you.
sources · Statista · Microsoft · McKinsey · PwC AI Jobs Barometer 2025/2026 · Lightcast
technique2 Sep 20263 min
Short beats long: why padding your prompt makes it worse
"Make it really good and detailed and professional" contains zero instructions. Here is what actually carries information — and why we built a whole scoring engine that refuses to reward length.
Take two directions for the same tool. One: "make a rent splitter that is really good, modern, clean, professional, user-friendly and impressive." Two: "split rent by room size; inputs: total rent, shared bills, each room's share; output: what each person pays this month, working shown."
The first is longer and says nothing. The second is shorter and says everything. A model can only act on distinct specification elements — what happens, to what, under which rule, shown how. Adjectives are not elements. Padding is not context. This is why Makery's articulation scoring counts elements and penalises filler: a short precise direction must beat a long padded one, or the metric teaches the wrong thing.
The working recipe: name the thing (a splitter, not "something great") · name the rule (by room size) · name the inputs and the output · name what done looks like (working shown). Then stop typing.
If your next prompt is longer than your last one and the output did not get closer, you added words, not information. Cut it in half and say the one thing that was missing.
from the makery scoring doctrine · nothing rewards length
the evidence1 Sep 20264 min
Cognitive debt: how AI quietly lowers your value — and how it raises it
The same tool measurably makes some workers sharper and others duller. The studies on both sides are real. The dividing line is not the tool — it is who is doing the thinking.
MIT Media Lab wired writers to EEG and had some of them write with an LLM. The assisted group showed lower cognitive engagement during the work — and when later asked to perform without the model, they did worse than people who had never used it. The authors call it cognitive debt: the thinking you skip today is a loan, and it comes due.
A 666-participant study (Gerlich 2025) found heavy AI use associated with weaker critical thinking, mediated by cognitive offloading — you stop evaluating because the machine sounds sure. And output homogenises: when everyone routes their ideas through the same models, everyone starts sounding the same (Doshi & Hauser 2024).
Yet the wage data points the other way: a 62% advertised premium for AI skills, productivity up 4× in exposed industries. Both bodies of evidence are right, because they describe two different ways of using the same tool. In complex work, AI raises preparedness and control — the human directs, the machine drafts. In passive use, the human drifts from conceptual engagement to surface-level acceptance — supervising output instead of directing it — and the skill atrophies.
AI raises your value when you direct it and keep thinking. It lowers your value when you hand it your thinking.
Every Makery mechanic exists on the right side of that line: you produce before the model responds, you reject outputs with reasons, you verify claims outside the model, and the record keeps score of all of it. Not because it is virtuous — because the alternative measurably makes you worse.
sources · MIT Media Lab · Gerlich 2025 · Doshi & Hauser 2024 · PwC
in the ovencoming
Show the recipe: what a hiring manager actually reads
Ninety seconds, four numbers, one artifact. What survives contact with a real hiring decision — and what never gets read at all.
in the ovencoming
Keep, cut, change: the instruction that saves your best work
"Try again" throws away the parts you liked. The three-word structure that keeps them — and why rerolling is not iterating.
in the ovencoming
From the dough up: your first week of directing
What the first seven days on Makery look like — first prompts, first rejection with a reason, first thing a real person uses.
fresh from the oven · weekly
Get the Recipes in your inbox.
One short email a week — a sharper way to direct AI, a prompt worth stealing, and what the best
makers pulled off. No spam, and you can leave whenever you like.
no spam · one email a week · unsubscribe anytime
updates · what shipped
Out of the oven.
3 Sep 2026
The journey grows to eight Stopsnew
New audit modules join the spine — catch the hallucination, what AI cannot do, the right model
for the job — and a whole new Stop, the Long Bake: feed it what it needs, use it right, keep your
own edge, take it anywhere. Stop 8, Connect Everything (18+ · the Founder plan unlocks its tools, never its certificate), takes your build to the real stack:
Supabase, n8n, Vercel, MCP. And the pantry: the journey is free forever — you pay only to generate
assets of your own.
3 Sep 2026
The Studio becomes a guided journey
Every Stop got a module, an example, how it works, a quiz, and hands-on practice in the live playground — skinned to your interest track. Plus: the "Why now" numbers on the front page, and this Recipes page.
2 Sep 2026
The dashboard preview — all eight surfaces
Home, Studio, The Counter, Stream, Journey, Coach, Portfolio and Get found — the full logged-in product, faithful to the doctrine: bands not percentages, process-ranked showcase, a coach that only asks.
2 Sep 2026
Ten questions, one FAQ — and a face for the bakery
The FAQ grew to ten straight answers in two columns, the kawaii-bread mascot took its place in the nav, and Log in moved next to Start free.
1 Sep 2026
The front door opens
makeryapp.xyz went live: the landing page, seven-step onboarding, the business inquiry form, and the weekly newsletter — all running on the new design system, dark-native, glow not shadow.