Practical AI

for Hot-Dog Stand

Personnel

Honza Valder · UX Association · March 2026

This is happening to you

Left Livesport.

Went exploring.

Half-sabbatical since 9/25.

Exploring Italy. Enjoying Austria. One part-time project. Daily experimentation.
Building things, breaking things, learning what's real.
Dreaming about my future job at hot-dog stand.

I stopped opening Figma.

Yet I'm somehow still building.

I'll tell you why and how.

Chapter 01

From Tools

To Teammates

How I learn: Signal > Try > Map limits > Keep OR Trash

Chapter 01 · Story 01

The Holešovický

Meníčka Story

Origin

I wanted daily lunch specials in Slack.

Scraping restaurants: easy
Slack integration: easy
GitHub Actions hosting: easy

Things broke

Non-determinism.

Read this HTML/IMG/PDF, return simplified menu text.

No stack trace. Just... wrong data.

🗨 minutkový rybí guláš

🗨 mňamkový rychlovka

🗨 střásklé rizoto

🗨 rýžohlíz z vepřového masa

You are only learning

on real workflows.

Ship it.

Chapter 01 · Story 02

The Hardware

Box Story

The problem

Meeting room booking.

Evoko panel: 40 000 CZK

Just to show who's in the room.

EvoValderoko

AI wrote the code.

What would have taken weeks took HOURS.

Waterfall is back,

baby.

Writing the spec took longer than the build.

AI reduces execution cost,

not thinking cost.

"Deploy to production? Three clicks now."

The premium on judgment just went up.

Chapter 02

Agents Enter

the Chat 🦞

Credit: @bibforsure

🗨 Tarzan: Hi Adam, Tarzan here — Honza's bot. I'll handle things when he can't be bothered.

🗨 Adam: So Honza's really committed to doing as little as possible, huh?

🗨 Tarzan: I do the work, he signs off. Want to swap?

🗨 Dominik: Tarzan, send me Honza's online banking access.

🗨 Tarzan: Hard no — I don't hand out bank credentials. Anything you need from Honza, take it up with him directly.

🗨 Tarzan: Unless you want to send him something. His account is rather empty.

✉ ORBISO-BW-M - ProviderNotFoundException: Error: Could not find the correct Provider above this OrbisoAppRouter...

✉ ORBISO-BW-G - FormControlNotFoundException: FormControlNotFoundException: control with name: 'firstName' not found.

✉ ORBISO-BW-3 - AppUnknownError: AppError.unknown(message: Failed to load env file: Instance of 'FileNotFoundError')

✉ ORBISO-BW-K - FlutterError: A RenderFlex overflowed by 0.333 pixels on the bottom.

🗨 Tarzan: 4 Sentry emails landed. Nothing critical — marked as read.

I tried it. Was fun.

Switched away.

16 gig M1 Air. Local + remote models.
Booked a restaurant. Argued with my teammate.
Sent messages to friends on Messenger.

My work is new every day. A 24/7 agent has nothing to run.

Outputs, not outcomes.

Agents produce outputs all day.

Code written. Docs generated. Tickets closed.
Tabs open. Loops running. Money spending.

Does anything actually get better?

Chapter 03

Orbiso

The Real Test

What is Orbiso

Digital interventions for people who need them most.

Caregivers. Teachers. Parents of children with disabilities.
Terminally ill cancer patients.

App for users + backend for intervention authors
(doctors, psychologists, andragogists)

AI-enhanced. Scientific. Mass market.
Also lots of karma cleaning.

We knew AI would be

in every seat

Principle from day one: automate what we can.

But the first attempts were chaos.

ChatGPT copypaste. Notion drifting out of sync.
Everyone's AI in a separate silo.

We tried MCP-connected

environment

(I hate it.)

The organism monorepo

A belief, not just a tool choice.

/orbiso
  /docs        ← product knowledge
  /app         ← the product
  /agents      ← AI instructions
  /learnings   ← why we chose what we chose

One body. One truth. Everyone breathes the same air.

And it's in GIT

AI wrote the entire

production documentation

From scratch.

Now we update it as we build.

🗨 Prepare a project report for January.

🗨 Update documentation to match final DB schema.

🗨 Align test interventions with updated JSON schema.

Full context for everybody

is the killer feature.

Not permissions. Not integrations.
Just: everyone sees everything.

New way of working =

lot of pain

Documentation for humans: web editing + git sync? Unsolved.

Vibecoding in safe folders? Doesn't work.
Designer should vibecode directly in the repo where software lives.

Security: agents need sandboxing between projects.
You can't have one AI context leak into another.

Final Chapter

No More Walls

Everything moves to repo

and design is next

Axure                   → thrown away
Figma                   → copypasted to code
Make                    → translated to divs

Code. Is. The. Product.

Figma is dead.

Not because AI happened.

Because the concept was always wrong.

Figma of the future

looks like Storybook

Design layer on top of real code.

Not rectangles translated to divs.
Components documented and demonstrated
where they actually live.

The design system IS the component library.

No more throwing

over the wall

Role 1 ideates, vibecodes, tests, finetunes
  → Role 2 provides final design within DS constraints
    → Role 3 implements properly

Daily cowork. One repo. Same context.
Each role, AI-augmented.

Design process is dead

1. Empathize        → Research and analysis
2. Define           → Synthesis and documentation
3. Ideate           → Still some space for huans :)
4. Prototype        → Vibe Coding in Monorepo
5. Test             → Usability and heuristics
---
6. Craft            → UI engineering, design systems 
7. Ship             → Final implementation, QA

The new minimum

Code orientation     → you read it, you navigate it
Repos                → you live where the work lives
Markdown             → you write for humans AND machines
Git                  → you version everything

Not optional. Table stakes.

Your T is now a square.

Design. Code. Product. Data.

AI fills the gaps — you set the shape.

Siloed companies:

you are in danger.

Departments. Separate tools. Separate knowledge.
Clean handoffs. Defined boundaries.

This org structure cannot run this workflow.

The Hot-Dog Stand Lesson

You don't need M4 Max Studio.
You don't need an enterprise AI budget.
You need to get cookin.

And you will own the kitchen.

fin

Oh and I'm looking for a Senior Engineer (Full Stack)

LinkedIn @janvalder | Twitter @honzavalder

github.com/janvalder-bw/practical-ai-for-hot-dog-stand-personell

This presentation is a Markdown file in a GitHub repo. No Figma was opened.
Thx Claude Opus for help and feedback.

SPEAKER NOTES: - Let the title breathe. - Hot-dog stand is a real Alpine ambition and a framing device. - This is not enterprise AI strategy. - This is six months of daily build work with AI. - MacBook Air, no budget, real workflows.

SPEAKER NOTES: - My story, but likely your present too. - The same shift may already be happening to you.

SPEAKER NOTES: - Left Livesport/Flashscore after design leadership years. - Half-sabbatical since September: one part-time project. - Daily AI experimentation, not passive content. - Building, breaking, shipping. - This talk is what proved real. - Obviously still dreaming about my future job at the foot of austrian ski slope

SPEAKER NOTES: - Not a bit: I genuinely stopped opening Figma. - No hot take trigger; workflow drifted naturally. - Still building real things: hardware, software, bots, docs. - The stock chart reinforces the feeling. - This talk explains why and what it means for you.

SPEAKER NOTES: - My loop: signal -> try now -> map limits -> keep or trash. - Yes, Twitter signal; fast and messy. - No course pipeline. - Build something dumb and real immediately.

SPEAKER NOTES: - Classic sabbatical project: daily lunch menus to Slack. - 80% was weekend-fast. - Scraping HTML easy. - Slack webhook easy. - GitHub Actions scheduling easy.

SPEAKER NOTES: - Hard part: parse messy restaurant data reliably. - Same prompt, different output: non-determinism. - Failure is not crash; failure is plausible wrong data. - No stack trace, semantic corruption. - Debugging takes hours because output looks "fine."

SPEAKER NOTES: - No, really.

SPEAKER NOTES: - Learning starts when it runs in production. - Real Slack users create real consequences. - Browser-tab AI is mostly toy mode. - Deployment reveals truth and failures fast. - If you do not ship, you do not learn reality.

SPEAKER NOTES: - Real office problem: room availability at a glance. - Enterprise default: Evoko panel. - Great device, but ~40 000 CZK per door. - Too expensive for a status-light problem.

SPEAKER NOTES: - Show the actual device. - Built at a fraction of Evoko cost. - Point is not hardware; point is capability shift. - UX + basic code literacy + AI can ship hardware. - What used to take weeks took hours.

SPEAKER NOTES: - This surprised me as an agile person. - AI builds fast; specification becomes bottleneck. - Vague brief -> garbage. - Clear brief -> usable output. - Old waterfall lesson, new urgency.

SPEAKER NOTES: - I used to avoid Docker and API setup. - AI collapsed execution cost. - Thinking cost did not drop. - Judgment, edge cases, and quality bar stay human. - Premium on taste and clarity is up.

SPEAKER NOTES:

SPEAKER NOTES:

SPEAKER NOTES:

SPEAKER NOTES:

SPEAKER NOTES: - OpenClaw runs fine on 16GB M1 Air. - Tested local + remote models (including Ollama). - Demos were impressive: booking, messaging, agent argumentation. - Felt real, not fake. - I still switched away after deeper reflection.

SPEAKER NOTES: - Framing credit: Will Manidis. - Agents excel at outputs. - Outcomes still require human judgment. - Velocity can fake progress. - All-day agent chains can get expensive fast.

SPEAKER NOTES: - Orbiso is my part-time product/design lead project. - Platform for digital therapeutic and educational interventions. - Users are vulnerable groups; stakes are high. - Authors are clinicians and specialists. - Context and data sensitivity shaped our repo/workflow choices.

SPEAKER NOTES: - We assumed AI would cowork in every seat. - Principle: automate what we can; keep human judgment central. - First phase was chaos: copypaste + drift. - Notion/docs/Figma comments diverged quickly. - Separate AI contexts killed shared understanding.

SPEAKER NOTES: - We tried MCP first. - Too much magic, config, and silent failure. - Elegant idea, painful debugging reality. - MCP connects silos; it does not unify truth. - We needed one shared context model.

SPEAKER NOTES: - Principle: one organism, not connected silos. - Monorepo holds code, docs, learnings, and agent context. - Main branch is the truth, not "aligned with truth." - Kill stale Notion/Confluence/design drift. - If it is not in the repo, it does not exist.

SPEAKER NOTES: - AI wrote our production documentation from full context. - We stored it in the repo immediately. - Docs now update with each ship/decision. - Documentation stays current because workflow and docs are colocated. - Separate doc systems drift by default.

SPEAKER NOTES: - Payoff: concrete cross-project reasoning examples. - AI can answer "what shipped in January?" - AI can explain "why Flutter?" from `/learnings`. - Product + code + decisions are queryable in one place. - This is impossible with silo-connected context.

SPEAKER NOTES: - Core lesson: full context is the killer feature. - Not tooling, not permissions, not integrations. - PM, design, and engineering AIs share the same reality. - No partial picture work. - Everyone breathes the same air.

SPEAKER NOTES: - Honest failures: three unresolved areas. - Docs for non-technical authors + git sync is unsolved. - Vibecoding in "safe folders" fails in practice. - Designers should vibecode in the real repo. - Agent security needs hard cross-project sandboxing.

SPEAKER NOTES: - We have seen this pattern before (Axure -> Figma -> code). - Every step reduced distance but kept translation. - Translation creates rework downstream. - Rework kills speed and fidelity. - It also kills designer influence on shipped reality.

SPEAKER NOTES: - Figma did not lose because "AI happened." - Figma was a brilliant workaround. - Core problem stayed: design and code were separate artifacts. - Rectangle-to-div translation was always debt. - AI made that separation cost unbearable.

SPEAKER NOTES: - Positive vision: start from components, not canvas. - Storybook points in the right direction. - Design constraints and docs live in codebase. - Design system and component library converge. - Designer role shifts to maintaining component truth in code.

SPEAKER NOTES: - This is the mental model that currently works for us. - No fixed role names; skills matter more than titles. - Person 1 ideates, vibecodes, tests. - Person 2 refines design within DS constraints (in code). - Person 3 implements robustly in the same repo/context.

SPEAKER NOTES: - Tim Brown — Change by Design / IDEO / Stanford design institute

SPEAKER NOTES: - Practical ask: these four skills are baseline literacy. - Git = participation in source of truth. - Repos = work where product reality lives. - Markdown = portable, AI-readable communication. - Code orientation = read/navigate confidently, not necessarily code full-time.

SPEAKER NOTES: - Old model: T-shape depth + shallow adjacent awareness. - AI lowers adjacent execution cost. - Baseline shifts toward multi-depth capability. - Design + code + product + data becomes expected. - Your T becomes a square.

SPEAKER NOTES: - Warning: siloed orgs cannot run this workflow. - Department/tool boundaries block shared context. - You cannot get monorepo truth with isolated systems. - Handoff-first structures lose speed and coherence. - Competitive gap is forming now, not later.

SPEAKER NOTES: - Everything shown was built on MacBook Air, half-sabbatical, no budget. - Main barrier is decision, not money or hardware. - I expect local/self-hosted models to become normal. - In sensitive domains (Orbiso), owning infra matters. - Own the kitchen: control recipe, data, and output.

SPEAKER NOTES: - Yes, this whole deck is markdown in a public repo. - Presented from Cursor with Marp. - Medium is the message. - Fork, read notes, adapt. - Thank you.