CRAFT: a method for using AI without losing your edge

My last post diagnosed the disease: AI quietly switched off the forced feedback loop that reality used to run for you — the compiler error, the failing test, the blank stare of a reviewer. The pressure that made you learn is gone, and the easy path now wears the mask of success. This post is the prescription.

Not a loose list of tips — those evaporate the moment you're tired. A named canon you can recall under pressure, the way SOLID works for object-oriented design. I call it CRAFT: five principles, one letter each, sitting under a single axiom. The name is the whole point — coding stays a craft, and AI is just a tool in the hands of someone who can still do the work without it.

The axiom: the moving line

Before the five letters comes one rule that governs them all — the way "single responsibility" sits at the heart of SOLID:

AI does what you have already mastered. You do what you are still learning. The line moves as you grow.

Hand the mastered part to AI and you go faster — that's pure win. Hand the part you're still forming to AI and you simply never form it; you pay in skill for a speed you didn't even need. Every principle below is just a way to stay honest about where that line sits for you today.

CRAFT — five principles, one acronym

Read top to bottom, the letters are a workflow: challenge yourself first, review what AI gives you with hostility, ask why instead of what, manufacture friction on purpose, and test that you genuinely understand.

C — Challenge First

Principle. Before you open the AI, attempt the problem yourself for 5–15 minutes. The first solution on the page should be yours — even a clumsy one. AI comes in to unblock you, never to start for you.

Why. The struggle is what builds the mental model — the neurons fire while you're stuck, not while you're reading a finished answer. This is exactly the friction reality used to impose for free; now you have to schedule it yourself.

R — Review Adversarially

Principle. You don't let AI review your code — you review the AI's. Read every generated line hunting for the bug, the edge case, the security hole.

Why. Judgement is the most durable skill in engineering and the only one that grows more valuable as the models get stronger. Reviewing AI sharpens it and catches the confident mistakes models love to make. Accepting a diff unread erodes both at once.

A — Ask Why, Not What

Principle. Use AI as a tutor, not a ghostwriter. Ask for the reasoning, the alternatives, the trade-offs — not just the finished snippet.

Why. "What" hands you a fish; "why" hands you the rod. "Give me the code" leaves nothing behind. "Why this approach over that one, and what breaks on empty input?" leaves you a level smarter than you were.

F — Friction on Purpose

Principle. Reality no longer forces you to strain, so add the resistance back yourself. Do regular AI-off reps — build something from scratch with no assistant. Retype code instead of pasting it. Close the AI and reproduce what it just taught you.

Why. Skill sticks only through desirable difficulty. A copy-paste leaves no trace in memory; retyping the same lines slowly does. Friction feels like a tax on speed — it's actually the price of keeping the skill at all.

T — Test Your Understanding

Principle. The master gate is a single question: could I explain this, line by line, to another person? If not, you don't own the code and you don't ship it. Predict the output before you run it; break it on purpose and fix it back.

Why. "It works" is not "I understand" — that gap is the illusion of competence, and it's exactly where deskilling hides. Testing your own understanding is the cheapest insurance against waking up one day unable to work without the tool.

CRAFT at a glance

Self-check: growing or decaying?

You don't need a study to know which way you're trending. You're decaying if:

• You can't write a basic version of the thing without AI.
• You paste code you couldn't explain if someone asked.
• You feel anxious at the thought of working with the assistant switched off.
• You've stopped reading AI's output and just accept it.
• You can't predict what a snippet does before you run it.

You're growing if:

• AI makes you faster at things you genuinely understand.
• You routinely catch the model's mistakes.
• You can still work without it — slower, but you can.
• You use it to learn a new concept, then reproduce that concept unaided.
• You decide consciously what to hand over and what to keep.

If you take only one letter

Take T. The explainability gate — could I explain this, line by line? — is the simplest habit in the whole method, and on its own it catches roughly 80% of the decay. Everything else in CRAFT just makes it easier to pass that gate honestly.

The bottom line

Nothing fundamental changed about learning. Conscious understanding was always what mattered; cramming was always a poor substitute. What changed is that the environment stopped enforcing that difference for you. CRAFT is simply a way to enforce it on yourself.

The rules of the game are the same. The game just got more treacherous — the easy path now wears the mask of success.

If you haven't read the diagnosis behind all this, start with the previous post on AI and the lost feedback loop. That one was the disease; this one is the cure.