WaysConf 2026 · masterclass · 16 September, 14:55, ROOM-PM · 90 minutes Szymon Paluch — waysconf.com
Most AI demos stop at wow: a nice-looking output that falls apart the moment it meets a real requirement. In ninety minutes you will put a coding agent in a loop with a checker it is not allowed to argue with, hand it a design, and walk out with a website on the internet that it built and corrected itself.
You need two things. Neither of them is code.
| The design | The TURBINE file: view it in Figma, and download it from the workshop page as turbine.fig, one file the agent decodes by itself |
| The prompts | Eight things to say to your agent, with a copy button on each |
There is nothing of ours to clone, no script to run and no code to read. You start in an empty GitHub repository of your own. Everything else, the project, the program that checks the page and the page itself, the agent makes for itself. If a step runs out of time, a rescue pack from the workshop page lets you carry on from the next prompt.
Before the day: the setup page. Four free accounts, fifteen minutes, much easier at home than on conference wifi.
Almost everything people build on top of coding agents is one of two shapes, and both are things you say rather than things you install.
A loop — do this again until a condition holds. You define the exit condition; the agent keeps going until a program, not an opinion, says it is met. → prompt 05
A workflow — do these things, in this order, with a bar between each. You define the phases and what has to be true before the next one starts. → prompt 08
They nest: the workflow gets you from nothing to nearly-right, and a loop inside one of its phases closes the last gap. Seeing which one a piece of work wants is most of the skill, and it is the part that survives whichever tool you are using next year.
Prompt 03 asks the agent to build the thing that decides whether the page is right, before the page exists. Prompt 04 builds the page, and prompt 05 then forbids the agent from touching that checker ever again.
A loop is only worth building if something in it can say no without asking a language model for permission.
A model asked "is this good?" will say yes, warmly and at length, and it will sound the same whether the answer is true or not. A program asked the same question exits 0 or it does not.
The repository is for the speaker and for anyone curious afterwards. A participant never opens it.
| Path | What it is |
|---|---|
prompts/ |
The eight prompts, in markdown. The page participants use is generated from these. |
design-pack/ |
The download: three PNGs, tokens.json, content.md |
guideline/ |
The setup page, and the prompts page built from prompts/ |
design/ |
Tokens, the Figma build spec, the generated imagery and the prompts that made it |
figma-plugin/ |
A Figma plugin that constructs the TURBINE file from the same tokens the code uses |
src/, public/ |
The reference implementation — what the design pack is rendered from, and what is live |
checks/ |
The verifier for that reference site. Not part of the workshop: participants write their own in prompt 03. |
deck/ |
The slides, as markdown |
evidence/ |
What actually happened when this was built and tested. Start with INCIDENTS.md. |
Everything the slides assert was measured, and the thing that measured it is in here.
evidence/INCIDENTS.md— thirteen real failures hit while building this. Nine were failures of the verifier, not of the page. One run had nine gates producing 427 confident, correctly-formatted failures about somebody else's website.evidence/prompt-trial-codex.json— the prompt pack run against a real agent in a real empty folder. The checker prompt (prompt 04 then, prompt 03 now) produced a 296-line checker in 568 seconds, and it independently decided to treat axe's incomplete results as failures — the exact blind spot that took a person an afternoon to find by hand.evidence/determinism.md— the same build checked twice. Identical verdicts; individual Lighthouse samples fourteen points apart.evidence/dry-run-claude.md— a trial that did not run, recorded as a non-run rather than left as a gap.
- Automated accessibility tooling reaches roughly 30–40% of what WCAG actually requires. A green gate means no known defect, never accessible.
- A pixel diff detects that something changed. It has no idea whether the change was an improvement.
- Nothing here has an opinion about whether the design is any good.
- Prompt 07 uses a model to check a model. It catches what the deterministic checks cannot, and it is the least reliable part of the whole thing. A second opinion, not an oracle.
The most useful thing in this repository is the list of ways its own verifier was wrong.
Code MIT, design and copy CC BY 4.0, the three typefaces under the SIL Open Font License.
See LICENSE. TURBINE is fictional; artists, copy and imagery are invented.