Woran wir arbeiten

Error messages in dark mode sat at a contrast no one could read, and fourteen pages gave a screen reader no way into the page body. Every colour pair on the site now meets WCAG AA, checked by the test suite rather than by eye, and every page has a landmark to skip to.

148 fixed-scope services with published prices, and nine whole projects for the work that has to be scoped before it can be priced. Every service page shows what sellers actually charge rather than a number we invented.

Answer a few questions and watch the range narrow, with the working shown. Most people will not fill in a form to find out whether the answer is €3,000 or €30,000 — they leave and ask somebody who published a number.

Write once, send to the developers who do that kind of work, compare what comes back. No account needed to ask, and the AI assistant that helps write the brief is forbidden from inventing requirements you never mentioned.

Work a person must do and work an agent can do are different things. Kept apart in the data, because in a single list an agent ticks 'get an API key' as done without doing it — and a false tick stops anyone looking again.

Record a path through your product — add to basket, check out — as steps with what should happen at each. It is the context a coding agent cannot infer from the code, and the thing that tells you a change broke something.

Capture a full page, record flows, stream console errors and failed requests to your editor, and fill forms from your brand kit. Signing in to the website is what connects it — there is no key to paste.

All 148 services translated, with technology names deliberately left alone: a translated stack stops matching what a seller wrote and what a buyer searched for.

Quote requests are saved and visible to sellers, but the sending domain is not yet verified, so the confirmation email does not go out. The app says so rather than promising a reply. Waiting on DNS.

A brief currently reaches sellers whose listed skills match the project category, which is a blunt instrument. It should weigh what somebody has actually delivered and how they were rated for it.

The 148 are curated and live in the repository, which is why every page is prerendered. Community submissions arrive at runtime and need moderation, so they belong in the database and get merged in — with the popular ones promoted into the catalogue proper.

Flows, tasks and captures already travel to the editor. The map an agent needs — what exists, what it is for, what depends on it — is the part that still gets re-derived every session.

An auditor can tell you 92 pages are slow; FlowKy can hand the fix to the coding agent in your editor. The bridge: audit reports posted to your verified project, grouped into one task per CAUSE rather than per URL, each carrying the evidence string your agent can search the repo for — and a re-scan closes the tasks whose causes disappeared.

Today a buyer and a seller agree terms and settle between themselves. Taking payment would mean escrow, disputes and refunds — worth doing properly or not at all, and not yet decided.