The Real Skills Behind AI Work
AI has made producing code and documents cheap. What it hasn't made cheap is deciding what to make, knowing when it's done, and checking it's right. Those are now the skills that matter most.
Domain knowledge still counts, but its role has changed. You need less of it to produce work, and as much as ever to judge it.
Five skills
Judgement
What to build
AI can generate ideas on demand. The scarce skill is choosing which ones deserve your time.
Specification
What done looks like
A vague target gets a plausible answer to a slightly different question.
Organisation
How to get there
Breaking work into pieces the AI can handle, and holding the thread across them.
Verification
Whether you got there
AI output is fluent and can be confidently wrong. Catching that takes domain knowledge, which is where it now earns its keep.
Diligence
Owning the result
Taking responsibility for what you ship, and for how AI was used to make it.
Compared with Anthropic's AI Fluency framework
This list overlaps strongly with Anthropic's 4D framework, developed with Rick Dakan and Joseph Feller. Diligence comes directly from it. Two differences are worth keeping: judgement sits further upstream, and organisation is named explicitly.
| This list | 4D framework | Difference |
|---|---|---|
| Judgement | Delegation | Earlier: is the task worth doing at all? |
| Specification | Description | Narrower: focuses on defining done |
| Organisation | None | Partly in Delegation; not its own competency |
| Verification | Discernment | Almost the same |
| Diligence | Diligence | The same |
Why judgement matters more now
The cost of making something has fallen, but the cost of owning it hasn't. When building was slow, the effort filtered out weak ideas. That filter has gone, yet someone still has to read the document, review the code and maintain the tool.
A quick test before starting
- Who will use this?
- What changes if it doesn't exist?
If you can't answer both clearly, it probably isn't worth building yet.
Why people still don't start
Time was never the real blocker; it was the respectable excuse. Developers have always talked about the app they'd build if they had time. AI has removed most of that time, yet many still don't start. Sometimes the time spent talking about it would have been enough to build it.
The likely real reasons:
- An unbuilt idea is perfect. Starting exposes it to judgement, mediocrity or failure.
- “An incredible app” isn't a specification. Without a clear picture of done, there's no obvious first step.
- Talking feels like progress. Describing the idea gives some of the satisfaction of building it.
- AI removed the typing, not the deciding. Scope, trade-offs, finishing and shipping remain. They were always the parts people avoided.
What this means
Coding looked like the barrier, but judgement, specification and organisation were the real ones all along. AI has just made that obvious.
So the most useful training now may not be another framework course. It's helping people decide what's worth building, define done before they start, and check results critically.