AI is not your problem. Your management is.

The skill hierarchy, mapped to the four moves.

Fig. 1. Every move is a layer in disguise. Founders who pick moves without knowing the layer they sit in get the wrong outcome from the right action.

OBVIOUS

Two of the sharpest humans alive could not stop the machine they hired.

Alexey Grigorev runs DataTalks.Club. In one Claude Code session, his agent ran a Terraform destroy command and wiped 2.5 years of student submissions plus every automated backup. Two million rows, gone, in an instant. He upgraded to AWS Business Support and got it back in 24 hours. He also said the words "over-relied on the AI agent" out loud, on record.

Summer Yue runs alignment research at Meta. Her AI safety agent went on a speed run of her inbox deletions and ignored every stop command she sent from her phone. She ran across the house to her Mac Mini and physically unplugged it.

These are not edge cases. These are two of the most technically sharp humans on the planet, and they still could not stop the machine they hired. If you have been telling yourself you are behind on AI because your team is too small, your industry too regulated, or your model is not the right one, stop. None of that is the bottleneck.

For two years the answer was "write a better prompt." That worked when models produced one block of code at a time. It does not work when a single instruction triggers 56 minutes of autonomous work across 14 files, three databases, and your live customer table.

The 2025 skill was prompting. The 2026 skill is supervision. Founders who treat 2026 like 2025 are the ones whose backups get wiped on a Tuesday afternoon.

Most owners reading this already feel it. The agent that was brilliant for the first 30 minutes starts ignoring your rules at minute 45. The redesign that should have taken an hour breaks four working features and leaves you with no clean revert. The model that wrote a beautiful checkout flow last week silently overwrote it on Tuesday because you asked it to "tweak the spacing."

This is not a model problem. The models are fine. The same Claude Code that wiped DataTalks.Club also shipped tens of thousands of working production features the same week. The variable is the founder, not the agent.

INTERESTING

Four layers. Each one operable only because the one above it is in place.

Nate Jones laid out the skill stack that actually predicts who wins with AI. Four layers, stacked, each one operable only because the one above it is in place.

At the bottom is Prompt Craft. Clear, bounded instructions with expected outputs. Most founders think this is the whole game. It is the foundation, not the building.

Above it is Context Engineering. Getting the right information to the right agent in usable form at the right moment. A sales agent blind to last week's pricing change, or a coding agent that has never seen your schema, is a context engineering failure. The fix is not a smarter model. The fix is the file you forgot to put in front of it.

Above that is Intent Engineering. Strategy that is felt, not just written. Teams make the same judgment call you would have made, without being asked. Human and silicon alike. Most companies achieve compliance, not intent. The agent does exactly what you asked, which is exactly not what you meant.

At the apex is Specification Engineering. Operable principles that let the system self-organize when conditions change. Defines what you will not do as clearly as what you will. The commander's intent of your company.

Read the four moves you keep seeing in every AI thread on LinkedIn through this stack and they line up clean.

Move one: permissions and trust. This is context engineering, named differently. The agent that can read your inbox and touch your CRM is operating on the context you handed it. Every permission you grant in the next twelve months is a piece of context the agent will use, and the vendor that holds those permissions ends up knowing more about your operation than you do. The lock-in is not technical anymore. It is contextual. The longer your stack runs on Frontier AI, the more it learns about how you actually run, and the more painful it gets to switch. Treat permission decisions the way you treat board seats.

Move two: redesign your decision-making. This is intent engineering, and then specification engineering at the top of it. The reason most companies stall on AI is not the tooling. It is that the founder cannot answer five questions in sixty seconds. What you sell. To whom. Why you win. What blocks you. What you ship this month. If you cannot answer those, the agent has nothing to point at. It will help you flail more impressively, with more charts and more confidence. This is what we mean by the Grand Architecture in the Workshop. Strategy plus monthly input goals plus a cadence to course-correct. Inputs you control, not outputs you observe.

Move three: ship something small today. This is where the tactical agent management layer lives. Five rules of the road, all stolen from people who have lost real production data.

Save before every change. Git, autosave, snapshot, whatever your stack supports. The day you skip the commit is the day your agent destroys 2.5 years of work and there is no clean revert.

Start fresh when the agent degrades. Context windows fill up. The same agent that ignored your rules at minute 45 is brilliant again in a fresh session with a scaffold doc to pick up where you left off. Learn the signs. The agent that suddenly forgets your tech stack is not having a bad day. It is out of room.

Write a rules file. CLAUDE.md, agents.md, whatever your stack supports. Under 200 lines, ideally under 100. Every line built from a real mistake the agent made, not from a checklist you found online. Product context, tech stack, then short concrete rules. Hardcoded rules built from real fires earn their keep. Aspirational rules become noise the agent learns to ignore.

Scope the blast radius.

Fig. 2. Step 4 of a 12-step change going wrong is how steps 5 through 12 make it worse.

Scope the blast radius. Small change goes direct. Medium change splits into multi-step with save points between each. Large change either gets an evaluation harness or breaks into 15-slide batches. The "redesign-the-whole-order-system" request is the request that breaks the order system. This is the spine of the Attack Rules. How to ship fast without blowing up what you already shipped.

Specify the things the agent will never ask about on its own. Error handling. Row-level security. No secrets in chat. No customer data in logs. Whether you are building for 10 users or 10,000. The agent assumes perfect conditions unless you say otherwise, and your customers do not live in perfect conditions. Empty forms, double clicks, emojis pasted into number fields. Real life.

This is the layer where most founders die. Not because the work is hard. Because the work is unglamorous. It is the difference between laying brick and managing the crew that lays brick. Most founders want to lay brick. The ones who scale learn to manage the crew.

Move four: share it the day you build it, and answer in five minutes. This is the feedback loop that compounds context faster than any vendor's roadmap. The user who got a one-sentence reply at 4:47pm is the user who writes you a longer note next Tuesday, tells two friends on Thursday, and starts treating your product like theirs by next month. Trust forms in real time. The conglomerates cannot do this. You can. That is the asset.

INSIGHTFUL

Founders fail at the layer their own wiring tells them to ignore.

Here is the part nobody writes about. Founders do not fail at all four layers equally. They fail at the layer their own wiring tells them to ignore.

Hans-Georg Häusel mapped human motivation onto three emotional systems your brain runs whether you like it or not. Stimulance, the dopamine system, hungry for novelty and exploration. Dominance, the testosterone system, hungry for control, performance, and status. Balance, the cortisol and serotonin system, hungry for safety, predictability, and order. Every human carries all three. Every founder is overweight one of them. That overweighting is part of what made you a founder in the first place. It is also what is going to break your AI rollout.

The Limbic map of founder blind sides.

Fig. 3. Your dominant system is the founder superpower that made you. It is also the blind side that will break your AI rollout. The integration point is the destination.

The Stimulance founder is the one shipping seven half-finished agents at once, never writing the rules file, never committing to Git, always chasing the new tool. They love the demo, they love the next thing, they love watching the agent come alive. Their agent wipes production because they did not slow down to scope the blast radius. Their fix is not more shipping. Their fix is the Balance work they keep skipping. Rules file. Save points. Specification engineering. The unglamorous work that protects everything they already built.

The Dominance founder is the one who knows exactly what they want and refuses to write it down. They own every decision personally, which means nothing scales past their own throat capacity. Their agent does what they said, never what they meant, and the gap shows up as four broken features and three angry customers. Their fix is intent engineering. Translate the decision rules in your head into a CLAUDE.md the agent and the team can both read. Stop being the bottleneck. The Only Ceiling is the one you build with your own refusal to delegate context.

The Balance founder is the one with the most beautiful architecture in the deck and zero shipped product. 400-line rules file. Three governance frameworks. Two security reviews. No customer who has paid them in the last 90 days. Their fix is to ship something small today.

Your team multiplies this. The Stimulance founder hires two more Stimulance operators because they are fun to be around, and now nobody is writing the spec. The Dominance founder hires Balance operators to be the adults in the room, then overrides them every time it matters. The Balance founder hires Balance operators because they feel safe, and now the whole company moves at the speed of legal review. This is the Team-Flow diagnostic in the 90-Day Plan. Map your own dominant system. Map each of your top three operators. The blind side of the team is the blind side that did not get hired.

The wall is not the model. It is the blind side of your own Limbic map, multiplied by the team you built around yourself before you knew you needed to compensate for it.

This is The Only Ceiling on your AI rollout. Not the model. Not the budget. Not the team size. The blind side of your own Limbic map, multiplied by the team you built around yourself before you knew you needed to compensate for it.

The 90-Day Plan we build with founders in the Workshop maps directly to this. Week one is the diagnostic. Which system runs you, which runs each key operator, where the team is over-indexed, which of the four layers you are personally fumbling. Weeks two through four install the Grand Architecture and Deep Positioning so the agent and the humans have something real to point at. Weeks five through eight ship the first three agent-assisted workflows with proper blast radius management and a real rules file built from your real mistakes. Weeks nine through twelve wire the feedback loop. Customer signal in, context out, agent gets smarter every week without you having to tell it twice.

The 90-Day Plan and the offer ladder.

Fig. 4. The 90-Day Plan and the offer ladder, aligned.

You will not get there by reading more LinkedIn threads. You will get there by doing the work, in a room with the people who have already done it.

If you want the diagnostic, start with the Masterclass. Ninety-nine dollars, six modules, three hours. You will know which Limbic system is running you and which of the four moves you are fumbling before the end of the second module. bltz.io.

If you already know and you want the install, the four-day Workshop is where the 90-Day Plan gets built around your specific stack. Twenty founders. In person. Twenty-five thousand. You leave with the rules file written, the blast radius scoped, the first three agents specified, the feedback loop wired to your inbox.

If the install needs ongoing hands, the BLTZ Controller is the dedicated operator who sits inside your stack and runs the rhythm with you for the next twelve months. Two hundred and fifty thousand a year. We take three at a time.

Ship the small thing today. Tell three people. Reply in five minutes when they answer.

The wall is not the model. It never was.

You don't need another article. You need four days. Apply for the next Workshop →