← Portfolio
Short case study · How I work

I do not just prompt AI. I design the system around it.

AI is good at making things. But ask it the same thing twice and you can get two different answers. That does not work when you are building real products.

5 skillsReusable
3 projectsOwn memory each
Every stepChecked
4 productsShipped
Any projectSame system

The output is not the skill. The system that repeats it is. Design it once. Run it on any project.

01The problemPhase 1 of 5

AI forgets. A system does not.

Every new session starts almost from zero. So the same work gets explained again, and the same mistakes come back.

Without MondayExplain the project ✕ TuesdayExplain it again ✕ WednesdaySame mistake returns
With MondayDecide it once → Written downWhere the work happens → Every day afterCarries on from there

It does not know your product

Your colours, your file names, last week's decision. None of it carries over.

It answers differently

Same question, different answer. A new screen will not match the last twenty.

It cannot grow

What works once, by luck, cannot be handed to a second project or a second person.

A better prompt does not fix any of this. A structure does.

02The approachPhase 2 of 5

I stopped writing prompts.

A prompt gets you one good answer. I wanted the same good answer every time, on every project.

A prompt

What to do, once

You ask. You get something. You judge it yourself. Tomorrow you ask again and explain everything again.

ask → answer
ask → different answer
ask → different again
A system

How the work is done, every time

The rules are written down. The project remembers itself. The result is checked before it counts as finished.

rules + memory → work → check → same answer
… on any project
… by anyone

Do not make AI remember everything. Build something that tells it what it needs to know.

The idea the whole system rests on
03The systemPhase 3 of 5

Five parts. Each one written down.

The rules and the project memory feed the work. Nothing reaches the output until it has been checked.

SkillTHE RULES MemoryWHAT WE DECIDED SourceA SCREEN OR BRIEF Work AI BUILDS IT Check DOES IT HOLD UP Output READY TO USE rules context input result passes fails · go again
Nothing reaches the output unchecked. The skill says how the job is done. The memory says what this project already decided. If the check fails, the work goes round again instead of shipping.

What is actually written down

figma-recreateHTML screen → Figma pixelDesign work byopmProduct team webdev-assistantWebsite changes sikshaTeaches me to code

Design work

Its own memory
  • 00_START_HERE
  • 01_CONVENTIONS
  • 02_MEMORY_SYSTEM
  • 03_PROJECTS_INDEX
  • PROJECT_TEMPLATE

Content tooling

Its own memory
  • 01_Memory
  • 02_Skills
  • 03_Agents
  • 05_Assets
  • 06_Tools

This portfolio

Its own memory
  • 01_working_rules
  • 03_design_system
  • 04_copy_style
  • 05_decision_log
  • 06_verification

Same shape every time: the rules, the decisions already made, and how to check the work. A new project gets the same three, filled in for itself.

04It scalesPhase 4 of 5

The project changes. The system stays.

Three very different projects run on the same skills and the same checks. Only the memory is different.

Design work+ ITS OWN MEMORY Content tooling+ ITS OWN MEMORY This portfolio+ ITS OWN MEMORY Same skills Same checks WRITTEN ONCE Its own product Its own product Its own product
The middle column is the part I built. Add a fourth project and it plugs into the same column. Nothing in the middle has to be rewritten.
5Skills, reusable anywhere
3Projects running on them
1Way of checking the work
+1What a new project costs to add

One example, and the mistakes it now prevents

A skill that turns a finished HTML screen into a real Figma frame. Editable, with Auto Layout, not a picture.

Source

HTML screen

One file. Every screen and modal.

The skill

Read and rebuild

Finds the real source, copies the exact spacing and colour.

Result

Figma frame

Real Auto Layout. Text that wraps.

White bars appearNew frames come filled with white. Clear every one, then sweep the whole tree.
Text looks shortIt is not short, it is cut off. Set the width before the wrap, and in that order.
Commas vanishThe plugin has no Indian number format. Write the formatter by hand.

Each of these cost a day once. They are now rules inside the skill, so they cannot cost a second one, on this project or the next.

05What shippedPhase 5 of 5

A system is only worth what it ships.

These are products people use, not AI experiments.

The subscription page live on the Markets Mojo site. Subscription pageLive in production. The Post Builder tool, showing inputs, a live creative preview and platform copy. Post BuilderIn use. 26 templates, 5 brands. Bunny, the Android client manager, showing a home screen of tax deadlines. BunnyIdea to Play Store. 39 screens. Reminder Bunny, turning a spoken sentence into a structured reminder. Reminder BunnyAndroid and iOS. In daily use.

This portfolio too. Nine pages, no framework.

I stopped asking what prompt to write. I ask how the work should go every time, then design the system around that answer.

What changed

AI is one part

It builds. It does not decide what is worth building, or whether the result is good. That is still the design work.

It needs maintaining

A rule goes stale when the product moves. The system only stays useful if it gets updated when something breaks.

Narrow on purpose

Each skill knows one job well. A general one would guess, and guessing is what caused the mistakes in the first place.

← More work Post Builder case study Subscription page case study