Notes on designing with AI

Writing

6 parts of the Design by AI series, 3 posts and talks, and a prototype to try.

Design by AI

A practical series on building products with AI. AI helps build. Designers make the calls.

6 parts · since October 2026 · latest November 4, 2026

Product context

PRD, transcripts, prototypes, benchmarks

Skill

Rules the AI reads first.

UX layer

Claude in Figma

Builds flows from design-system components.
Opus 5.5 via MCP

Worklog

Decisions, sources, open questions. Context outlives the session.

Review and iterate

Designers, PMs and business heads review every flow.
people close the loop

Every round sharpens the skill’s UX layer.
01October 25, 2026 Using AI to design AI. “I use AI to design AI products. The useful part is not the model.” Read on LinkedIn (opens in a new tab)
Fixing the output

Patch each screen. The next one breaks the same way.

123 patches

Fixing the system

Fix it once in the checklist. Every screen after comes out right.

Checklist

02October 30, 2026 Fix the system, not the output. “Don't just fix the output. Fix what the next version starts from.” Read on LinkedIn (opens in a new tab)
1Check

Every path through the flow.

2Play

The sheet hides the title.

Welcome back

Title hidden

3Fix

Lower the sheet. The title clears.

Welcome back

03October 31, 2026 Check, play, fix. “How I audit my prototypes now.” Read on LinkedIn (opens in a new tab)
Where decisions get made

Each layer is reviewed before the next one builds on it.

Decisions made on purpose
1reviewed Input PRD, calls, Slack
2reviewed Define Problem, users, goals, KPIs
3reviewed Benchmarks What good looks like
4reviewed Flows + IA Moves and the bigger picture
5reviewed UI What it feels like
6reviewed Code If needed
One prompt skips all of this. The model fills the gaps with whatever it already knows.

04November 2, 2026 Step, then step. “Why do you still design in Figma when AI can write the code directly?” Read on LinkedIn (opens in a new tab)
One path first

Define together. Refine one path. Then expand.

1Define

Agree on the problem, who it’s for and the outcome.

The problem
Who it’s for
The outcome

Everyone involved, not just design.

2Build one happy path

AI drafts the main journey first, not every screen.

3Refine together

I question it, give feedback and steer. AI iterates. Repeat.

You

AI
feedback next iteration v1v2v3

4Align

Review the refined path with the team before expanding.

5Expand

Then add alternative paths, edge cases and the remaining states.

One path first

Define together. Refine one path.
Then expand.

1Define

The problem
Who it’s for
The outcome

2Build one happy path

3Refine togetherv1v2v3

You feedback →← next iteration AI

4Align

5Expand
05November 3, 2026 One path first. “AI makes building faster. It also makes it easier to jump into solutions before the team agrees on the problem.” Read on LinkedIn (opens in a new tab)
One week of work

Mess in, structure out.

Documents

Video

Voice notes

Dashboard

one week
06November 4, 2026 One week of work with Claude Opus 5.5. “None of this is launched yet. All of it has been delivered.” Read on LinkedIn (opens in a new tab)

Posts and talks

Standalone pieces, outside the series.

All posts on LinkedIn

A car's camera view on its dashboard screen: a silver Mercedes drawn in 3D in a parking spot, ringed by glowing blue guide lines, under the warning Check surroundings!

PostNov 2026 My Mercedes has a camera view that looks like a game trailer. Beautiful, and the least useful view for parking in Cairo. Every hidden view costs memory, every switch costs attention: the fancy view gets the screenshot, the plain view gets the job done. Read on LinkedIn Talk · BrainsMingle AI SummitOct 2026 Success in AI products is people fixing it less, not using it more. Trust is built in steps, the AI belongs inside the task, and one home screen for everyone is the wrong home for everyone. Read on LinkedIn PostOct 2026 Approving a purchase in a chat isn't the merchant accepting it. Agentic checkout means keeping three things apart: what I authorized, what the merchant accepted, and what happened to my money. Read on LinkedIn

PrototypeAug 2026

AI native ≠ chat box.

AI should be built into the task, not slapped on as a chat. Here a group chat becomes a confirmed dinner plan: the AI reads the conversation, fills what it's sure of, asks about the one thing it isn't, and leaves once a person says yes.

Blue while AI actsEverything the AI filled in looks like the AI filled it in.
Ask onceOnly the field it isn't sure of asks for a tap.
Gone on confirmOnce a person commits, every trace of the AI leaves.

Try it yourself How it's built Read on LinkedIn

Blob animation by Ahmed Khaled

The prototype, start to finish

← Home