Lesson 6 of 12 · 3 min read

Why does designing with AI feel slower than it should?

Because most of the time goes to explaining, re-checking, and repairing, and those trace to three gaps: the tool lacks the context, nobody decided what to hand it, and the team can’t yet tell a confident answer from a correct one. Each gap has a fix to try.

Diagram: context and constraints feed an agent, whose work passes Draft, Review, Repair, and Approve, with judgment at Review and Approve.
The draft comes out fast. The time goes to the two places a person has to judge it.

The same ten frustrations keep coming back

Devansh Tank collected them in a LinkedIn post: explaining takes longer than designing, small edits break approved work, the output drifts from the design system, and prototypes pass for finished products.

I’ve had most of them. I think they come from three gaps.

Six of the ten frustrations, and the gap each one comes from

What it feels like

  • Explaining longer than designing
  • Drift from the design system
  • Changes to work already approved
  • Paying for failed attempts
  • A prototype taken for a product
  • Doubting your own ability

Where it comes from

  • Context: it was never told
  • Context: the rules weren’t readable
  • Delegation: no checkpoint before it changed
  • Delegation: no plan checked first
  • Literacy: nobody said what the prototype proves
  • Literacy: no way to tell right from sure

Context is the gap I hit most, and the cheapest to close

When I built After Send, a payments concept on simulated data, I first decided what an accounts payable specialist on the phone needs to see. Then AI built a working version in a day.

After Send’s payment record for Hillside Packaging: Processed by supplier beside Ready to match to invoice, and under them who acts next.
From my work After Send. The payment record, with what the specialist on the phone needs first at the top. See After Send

The explanation is long because it was never written down.

Drift is the same gap. This site’s written design rules drifted from its code three times before a check compared them. Now the rules are where a tool reads them.

Delegation means choosing where you’ll look, before it starts

I don’t wait for a finished prototype. I check the plan, one representative component, and the error handling before calling a flow done.

In a simulated run of Agent Review, my prototype of product design with AI agents, I told the agent to change nothing anyone could see. When a fix would have made a cancel link easier to find, it stopped and asked.

Literacy is the slowest of the three. A message saying the work is done is a claim, and a team learns to check it.

Some of the slowness is the work you skipped before you asked. The rest is checking, and checking is the job.

Check yourself

What did you last explain to an AI tool that is written down nowhere else?

That’s context living in a chat. Put it where the next session, and the next person, will read it first.

At what point do you check the tool’s work: during, or only at the end?

Checking only at the end is where approved work gets broken. One checkpoint on the plan saves much of the repair.

The idea: Devansh Tank, on the problems teams face designing with AI. The examples: After Send and how I built this site.