How I designwith AI.

How AI fits into my design work, from a fuzzy spec to a merged pull request.

Arunkumar AlagarsamyProduct Designer
spec-to-pr · one feature, end to endthe loop
$ claude run feature --from=spec
… reading the spec … 7 steps found
research.contextcontext.md, for me + Claude repo.artifactrepo-map.html prototype.buildrunning on localhost experience.itv1 → v3, gaps fixed edge.casesAI-imagined, now tests stakeholder.reviewfeedback on real links ship.prPR open, review requested
experience first · AI does the legwork
$ gh pr view
Why this matters now

When code gets cheap, user experience and taste become the last moat.

Faster

Solving the wrong problem faster

AI can ship the wrong thing in a day instead of a month. It's still the wrong thing.

Better

Solving the right problem

Knowing what people need and making it feel right. That's the part that can't be generated.

You've got to start with the customer experience and work backwards to the technology.

Steve Jobs · WWDC 1997
80%solving the right problem, the right experience
20%the build
01

Research with AI

02

Learn with Artifacts

03

Quick Prototype

04

Experience it

05

Edge cases & stress-test

06

Collect Feedback

07

Ship & measure

the data starts the next round
01Research with AI

Research warms up Claude Code, and me.

An agent doesn't know the nitty-gritty of my problem and can't one-shot a solution. I steer it with context.

Cartoon of a man in a context hat and context shirt holding a sign: I require contextevery AI agent, before warm up

Context inputs · I steer

Specwhat's asked vs. assumed
Datathe numbers behind it
Questionswhat I'd need to explain it
Unknownsgo deeper until clear
Answersunderstood, in my words
Trade-offswhat we give up, and why
▤context.mdshared
Problem
who it hurts, and when
Users
what they already do today
Data
what the numbers say
Limits
tech, time, business rules
Decided
answers, in my words
Trade-offs
what we give up, and why
Open
unknowns still being chased
grows with every round of questions

Warmed up

Claude Code

Starts from my problem, not a guess.

Me

Can explain it before I design it.

02Learn with Artifacts

Artifacts turn complex information into consumable, interactive diagrams.

Claude explains the code as a page I can click and play with. It warms me up, far faster than reading files cold.

BeforeWorking on a new repohow it's built, before I change anything
AfterBuilding my own featureproof I understand what I shipped
↘ click one to open it, live
03 · 04Prototype, then experience it

Creative ideas show up on an
interactive prototype, not a static screen.

Claude Code builds a rough working version. I use it like a user, and each gap I feel becomes the next idea. First Rank Pro's detail page, live:

1A list of checks, no statusWhich ones fail? Click each one, on every page.
2A status on every checkFails at a glance. But why? Still open each check.
3Quick SummaryEvery check and its value on one screen.
First Rank Prov1 · no statusv2 · status iconsv3 · Quick Summary✕
which checks fail?…but why does it fail?the whole page, in one look
05Edge cases & stress-test · ConvertKitty

AI imagines the edge cases I'd miss. Each one becomes a test.

I ask AI

What files could someone use to break this converter?

+Opus 5.5
AI-provided edge cases
a zip bomb: 300 MB hiding in 299 KB
a file named ../../evil.txt
a path straight to /tmp
10,001 files in one archive
a symlink to /etc/passwd
a ZIP renamed to .rar
encrypted, and no password

→ 9 hostile files, generated on every test run

One of them, as a test
evil-bomb.zip299 KB on disk
299 KB×1
limit ×100
299 KB300 MB of zeros
Refused
SUSPICIOUS_RATIO

“The archive expands far more than a normal archive (possible zip bomb).”

Caught from the packing list, before anything is unpacked.
05Edge cases & stress-test · Prompt to Blade

For UX, AI plays the user who pushes it off-brand.

I ask AI

How will people push this off-brand, or break it?

+Opus 5.5
AI-provided edge cases
“make the badge neon green #00ff00”
“add a FancyCard with a glow”
“set the heading to 13px”
an empty prompt
a 2,000-character prompt

→ each one gets a designed answer

One of them, as a test
make the badge neon green #00ff00
badge.color#00ff00→
positivenegativenoticeinformationneutralprimary
“a FancyCard”not one of the 11 element kinds “13px heading”small · medium · large an empty prompt400 empty_prompt 2,000 characterscut to 500
Impossible
enum · const · additionalProperties: false

An off-brand colour, a made-up component or a magic pixel can't even be written.

The schema decides, not the prompt.
05Edge cases & stress-test · Product Highlights

The edge case I never thought of: right-to-left.

I ask AI

What kind of storefront could break an embedded widget?

+Opus 5.5
AI-provided edge cases
!important on every class it uses
a transformed ancestor
a dark storefront
a missing anchor, bad selectors
a 320px page, right-to-left
a strict CSP
a 404, or broken JSON

→ 9 stress pages, each states its own pass test

One of them, live · the real widget
  • Nothing past the edgemeasuring…
  • Toast stays on screenmeasuring…
  • No clipped textmeasuring…

It mirrors itself: the CSS uses padding-inline-start and inset-inline-start, never left or right.

07Ship with the developers

Prompting all day doesn't earn a pull request. Knowing Git does.

github.com/arun-dev-des 1,072 contributions this year123 merged pull requests47 repos
01

Understand

An artifact of the module I'm changing.

02

Pull

The latest from main.

git pull
03

Branch

One branch per feature.

checkout -b
04

Build

With Claude Code, in the real repo.

claude
05

Test

On localhost, plus the stress suite.

npm run dev
06

Push

The branch, small commits.

git push
07

Pull request

Developers review it like any other.

gh pr create
Git Fundamentals certificate, GitMasteryGit FundamentalsGitMastery · 18 levels · Oct 2026
07Ship with the developers

AI can mess up how it writes a feature. So I read the code.

Not a vibe coder. Certified in the languages my developers write, so I can read every diff before I push it.

07Measure · data as context

I built the analytics dashboard, then hand AI the real numbers.

ConvertKitty · my own analytics · first 24 days
582conversions
92%succeeded
78%never touched the server
11+countries
Conversions per dayin the browseron the server
12 Oct232 on 28 Oct4 Nov
Where conversions came from
ChatGPT 50%Direct 39%Google 11%
conversions-24-days.csv582 rows

+Opus 5.5
What we did with it

Every converter and every format pair gets its own real page, so ChatGPT always has an exact page to link.

2,269working converter pages in the sitemap
07Measure · errors as context

Every failed conversion becomes a record. AI triages, I decide.

ConvertKitty · dashboard · Errors
32error records
20fixed and shipped
6expected refusals
ERR-0031Samsung Galaxy S26 RAW photos could not be converteddeployed
ERR-0028Server results lost when a phone reloads the tabdeployed
ERR-0022XCF → PNG on phones: the engine never finished downloadingdeployed
ERR-0021A 39,883-page PDF ran out of memory, reported as “damaged”deployed
ERR-0016Visitors could not find where to type an archive's passwordverified
ERR-0010Uploads over 1 GB were read for 30 minutes before refusalverified
ERR-0007A real .deb opened as emptyverified
ERR-0018JPG → PNG: the file is not a JPEGexpected
ERR-0016newrecommendedapprovedpr_opendeployedverified
Captured · 24 Oct

RAR → ZIP refused 4 times with ENCRYPTED, two iPhone visitors, same file. The refusal was right. So why four tries?

✳ Claude Code · “triage errors”

“The ENCRYPTED error did not lead the visitor to the password field.”

The password box was hidden behind Options: an error that was really a UX bug.
Me · in the dashboard
Approve the fixExpectedNeed more info
Shipped · PR #30

On ENCRYPTED, Options opens by itself with the password field. A wrong password asks again.

Deployed 25 Oct · verified 1 Nov, after 7 quiet days
Where the line is

AI widens what I can do. It doesn't decide what's worth doing.

AI does the legworkI make the calls
ResearchAnswers, sources, a first draft of the contextThe inputs: spec, data, trade-offs. I steer
ArtifactsTurns a codebase into interactive diagramsKnowing where my change will land
PrototypeWorking builds, in hours, many of themWhat feels wrong when I use it
Edge casesImagines the cases I would missA designed answer for each one
FeedbackTurns curated feedback into the next roundTalking to users, stakeholders, developers
ShipWrites the code in the real repoReading every diff, owning the pull request
MeasureTriages errors, crunches the numbersApproving fixes, deploying, what comes next
Photo: Matthew Yohe · CC BY-SA 3.0 · dithered

You've got to start with the customer experience and work backwards to the technology. You can't start with the technology and try to figure out where you're going to try to sell it.

Steve Jobs · WWDC 1997