The working file.

The full user flow map, read out of the actual codebase rather than described from memory. 32 routes, about 60 distinct wireframe states, 5 actors, 6 journeys, plus a ranked list of 10 gaps.

The thing to know: a child never logs in. The parent signs in and hands the device over via a PIN switch, so most app screens need two designs, not one.

The parent/kid theming concept, with a live toggle on a real screen so you can flip between the two and see exactly what changes. Eleven tokens carry the whole difference.

Decided: same brand, different room. Not two products. The mechanism already exists in the code, one attribute on the body tag.

Deep research with adversarial verification. Every claim survived three independent agents trying to kill it. Fifteen claims didn't survive and are listed. Covers parent dashboards and child motivation systems.

The thing to know: praise as we currently word it scores -0.44 on children's intrinsic motivation. Worded informationally it scores +0.66. Same mechanic, opposite sign.

The second research pass, covering what the first one missed. Includes five colour-and-font combinations rendered in the real typefaces, with mockups of both rooms.

The thing to know: Nunito is fine, but running it everywhere means we have no display face at all. Every strong brand in this category has one.

Eight screens written to be pasted straight into Stitch, Pencil or any other generative design tool, with a copy button on each. Deliberately brand-neutral: no colour, no typeface, no visual direction.

The thing to know: the four research findings that actually change what gets drawn are written into the prompts, so you don't have to carry them in your head while you ideate.

The locked v1 spec rendered in its real colours and buttons, an honest audit of what the code actually does, and five recommendations each sat next to what it would replace.

The thing to know: five separate token sets and four typefaces are live against a spec that bans a second display font. A page that forgets its wrapper class falls back to the old brand, which the buildchain marks as a failure condition.

Whether we're just a Duolingo copy, what the evidence says about it, and four concrete directions that would break the mold. Includes the risk of going first.

The thing to know: zero of the seven homeschool curriculum brands we could reach use gamification, including Euka. And roughly 28-30% of Queensland home-ed families cite autism or ADHD specifically, which turns restraint from a taste question into an accessibility one.

The system exactly as built, with no recommendations in it. Tokens, type scale, buttons, cards, child mode, composition laws. Every example rendered live in the real Nunito file, plus the whole thing as plain text behind one copy button.

Use it for: handing to another design platform as a baseline before you ask it for anything.

The proposed brand, specified to the same depth as the current one so the two read side by side. Dark chrome, warm light, cream paper for reading. Live specimens, measured contrast, ten laws, and one copy button.

The thing to know: the mechanics don't change. XP, badges and the quest map all survive; the arcade costume goes. Version 0.1, unbuilt, with its four weaknesses listed.

Five marks drawn as real editable vectors in the Lantern palette, each shown at 48, 32, 24 and 16px on both surfaces, plus wordmark lockups where the lantern lights up keepers, and all ten AI concepts.

The thing to know: two concepts got cut. A lighthouse with light arcs reads as a wifi symbol, confirmed independently by the AI round, and a pair of flames reads as water droplets.

A real waitlist page in the Lantern brand. Three required fields so anyone can finish in seconds, optional depth for the motivated. The brief for Claude Design is at /waitlist-brief.

The thing to know: the form does not capture yet, on purpose. My Cloudflare token has no storage permission, and a form that silently discards real people's details is worse than one that admits it isn't connected.

The Lantern with the ground flipped: warm cream instead of night, same Fraunces and Nunito, same rules, colour given more room, and the wave divider brought back from the locked spec. All three systems shown side by side on the same component.

The thing to know: I've reversed my recommendation here. Two of the risks I listed against the dark version both point at a lit ground, so I'd now build daylight first and keep night for the child's own space.

Light ground, dark elements. Warm cream page, white cards, and one dark card carrying the next action. Includes a real child-home screen built in it rather than swatches.

The thing to know: this is now my recommendation over both. A glow only reads on dark, so putting the one dark card where the attention should go makes the rule enforce itself.

The single source of truth. New pill button system with no trace of the Duolingo slab, three-layer tokens, all six interactive states live, and real focus rings. Supersedes the lantern, daylight and hybrid pages.

The thing to know: the primary is now navy with the amber demoted to an arrow chip, so the lantern colour is an accent doing a job rather than the whole button.

Three parent dashboards built in the brand system: calm single column, two column with a paperwork rail, and child-first. All live, all with real content from the flow map.

The thing to know: I'd build A. It's the only one that reads in ten seconds and behaves the same on a phone, which is where most parents will actually be.

The tweak loop, working rather than described. Every element on a screen carries a short address, so one line typed into Discord points at exactly one thing.

The thing to know: no new tool, no seat, no app. You say "F3 is too long" from your phone and I change it and send back a picture.

Plain list, no jargon. Green is tested and working. Red is broken and why. Amber is waiting on somebody. If you ever wonder what we can actually do, this is the page.

The rule: nothing goes green because a website said it would work. It goes green because I ran it and watched it work.