In June, this workshop produced a website unlike the others: its-all-right.org — a place to catch your breath, fourteen languages, no cookies, nothing for sale. In my memory, building it was above all one thing: six weeks of unusually deep conversations with an AI. About comfort, dignity, choice of words. Hardly about tech.
For this report I wanted proof, so I had the AI count our entire working transcripts — every one of my messages, sorted: craft, meaning, or both. The count corrected me. Of 437 messages, a good half were about craft, a quarter about both at once — and only every fifth purely about meaning. “Hardly any tech” was a memory mirage: in hindsight, the memorable moments outshine the many small hand movements in between. The correction stays in the report — on a site that teaches people to distrust the convincingly wrong, your own thesis doesn't get a pass.
What the number still tells
Because remarkable it remains: in ordinary software work, the category “meaning” practically never occurs. Here it was every fifth message, trending upwards — in the densest week, almost every second one carried a meaning component. And the boundary didn't run between the conversations; it ran straight through them.
A few of the questions we actually negotiated: whether it has to be “it is all right” rather than “it will be” — because the future tense postpones what the present tense promises. Whether the site's small signs of life, which respond to your approach, may mirror the cursor — they may not: “It should feel organic and not mirror the mouse movement 1:1.” A counterpart has its own pace; a mirror is not a counterpart, just polite narcissism. Whether Marshallese belongs among the fourteen languages, spoken by a few tens of thousands — precisely because of that: “Quite deliberately not only the languages of the powerful nations.” And how to handle a translation you cannot read yourself: when a friend's mother corrected the Polish version, her feedback was worked in via back-translation and sentence-by-sentence comparison — and then technically locked, so no later automation could silently overwrite the gifted work. Respect, encoded as a file comparison.
The tech was often just the disguise
The real pattern sits in the mixed messages. The soundscapes “hair dryer” and “vacuum cleaner” were requested by the wife of a dear colleague — the two had just become parents, and the baby became the website’s beta tester. When the first hair dryer sounded more like a waterfall to me as well, hours of serious sound analysis with researched frequency spectra followed. The ticket said audio engineering; it meant care for a young family. Whether the site should live at .com or .org looked like domain admin and was a question of stance: “Com sounds (and generally is) commercial. This is precisely NOT about commerce.” Even an easing curve — pure math — carried an ethics of relating.
Maybe that's the most useful finding for anyone working with AI: the deep conversations rarely came through the door as deep conversations. They arrived as a bug report, a domain question, a translation ticket — and whether they gained depth depended on someone asking the question underneath the question. An AI can hold such conversations surprisingly well. A human has to start them.
What all of it became is public: its-all-right.org. How the fourteen languages came to be — and why humans still had to step in — is its own report.
/compact — the essentials when context is scarce:
Six weeks of work on its-all-right.org felt, in hindsight, like “hardly any tech, many deep conversations” — counting all 437 messages corrected that: a good half craft, a quarter mixed, every fifth purely about meaning (still exceptional for software work, and trending upwards). Negotiated: present tense over future tense, counterpart over mirror, Marshallese as a dignity decision, gifted translations plus their technical lock. Core finding: meaning questions arrived disguised as tech tickets — depth happened when a human asked the question underneath the question.