Experience

Pick. A kid signs in (her parent is already verified) and picks one of five starter templates: clicker, timer, scorekeeper, drawing pad, reaction tester. Or blank canvas. The template renders in the Sandpack sandbox instantly. No waiting, no loading screen.

Talk. She types what she wants to change in plain language. "Make it bounce when I tap it." The AI streams its plan, asks a checkpoint question if the idea is fuzzy (exactly three choices, grade 5 reading level), and proposes a patch via submit_patch.

Watch. She watches the code come alive. The patch applies, Sandpack compiles, the new behavior works. If something breaks, the recovery flow takes her back without shame. Red is banned from the palette. The punch-warm tone handles warnings.

Ship. She taps ship. The app is bundled, uploaded, and a URL comes back. She texts the URL to a parent, a grandmother, or a classmate. That moment is the product.

Keyboarding considerations are real for the 8-year-old end of the cohort. The chat input is the primary entry point because kid keyboarding has been good enough for this long. TODO(product): voice input is on the roadmap as an accessibility surface for younger kids and kids with motor differences; a first integration lands in Stage 2 and is not in the Stage 1 critical path.

Verify. Parent starts by verifying age and identity through Persona (Stage 2 stack). Takes roughly 90 seconds. The consent record is stored with a dated artifact and linked to the kid account. No kid account exists without a verified parent.

Consent. Parent reads a plain-English summary of what Tekku collects, how long we keep it, who processes it, and how to delete it. Clicks consent. That click is the legal record. Grade 8 to 10 reading level on the consent page by policy.

Read. Every Sunday morning, parent gets an email: what the kid built this week, which concepts she learned, one suggestion for the weekend. Six to eight sentences, grade 10 reading level. Every concept claim carries an evidence line and a link to the real session snapshot.

Audit. At any moment parent can open the kid dashboard, see every session transcript, every flagged moderation event, every shipped app. Parent can delete any of it with one tap. No ticket in between.

The reason parents do not "just watch" is intentional. The product is designed to give the parent back the screen-time conversation, not to replace it with a surveillance feed. The weekly email is the primary surface. The on-demand audit is the safety net. The parent trusts the tool precisely because the parent does not have to hover.

A session is bounded by the 30-turn per-project cap (see lib/cost/caps.ts) and by the kid's attention. Most sessions are short. The median target is 20 to 40 minutes. TODO(validation): publish the actual distribution once enough founding-cohort sessions accumulate. Kid attention is the real constraint, not the cap.

Difficulty curves with the kid's prompts. The AI does not impose a level progression. If the kid asks for a harder feature, the AI builds the harder feature and the concept detector labels the concepts used. If the kid repeats a familiar pattern, the AI ships the repeat without forcing a new challenge. The difficulty walk is led by the kid, not by the tool.

Recovery from a break is the most carefully engineered sub-flow in the product. When code breaks, the AI does three things in sequence: translates the error into kid language, offers to revert to the last working checkpoint, and offers an alternative patch that avoids the broken shape. The kid never sees a red error modal. The checkpoint is always one tap back. Broken code is never a dead end.

The five templates (clicker, timer, scorekeeper, drawing pad, reaction tester) cover the most common kid first-build patterns. Each template has a per-template cached prompt that knows what good looks like for that app type (see the prompt architecture disclosure on the AI engine page). The template prompt cuts token cost on repeat sessions and lifts first-session quality at the same time.

Blank canvas is the escape hatch for the kid who walked in with an idea. No constraint, no starter code, a direct line to the AI. The blank canvas path has a slightly different system prompt: the AI asks a clarifying question early to ground the build, because without the template context it is easier for the AI to head the wrong direction.

The default screen shows both paths side by side. A kid who is staring has five templates to pick from. A kid who knows what they want can click straight into blank canvas. The cognitive load is low either way.

Desktop and laptop are the primary build surfaces because keyboarding is faster and Sandpack renders at comfortable size. Tablet is a first-class build surface; many founding-cohort kids build on an iPad in the kitchen or living room. Phone is deliberately not a primary build surface for the kid, because the keyboard and the preview compete for the same tiny screen.

Phone is the primary surface for two other jobs: parents reading the weekly email and opening the kid's shipped URL; grandparents and friends opening the shared URL. The ship output is mobile-first by design. The viral badge on every shipped app is sized for readability on a phone. The landing page at the referral URL is optimized for parent signup on a phone.

The split posture is why the product works as a household subscription. The kid builds where the build wants to happen (laptop or iPad in the afternoon) and the family experiences the output where family communication happens (group text on phones in the evening).

The Teko mascot is the product-long play that turns Tekku from a tool into a companion. Six interactivity layers are in the design: mascot reacts to ship events, mascot comments on concept coverage, mascot proposes next-ideas based on the kid's build history, mascot nudges when a session is stalling, mascot celebrates first-time-used concepts, mascot links out to other kid projects the kid might enjoy (curated, no direct messaging). The Tier 1 document is at docs/plans/tekku-gamification-plan-2026-04-22.md.

Rollout is timeline-dependent. Layers one and two are part of the Stage 2 surface. Layers three through six land with Stage 3. The order is driven by what we can ship safely and what adds the most weekly session retention without risking the parent trust posture. Teko does not ask for kid data the tool does not already have. Teko does not sell anything. Teko is the on-surface voice that makes the product feel alive.

The mascot work is tied to the creative doctrine. Kaze's Seven Laws (breath, tension, presence, honesty, memory, weight, silence) govern how Teko behaves. Teko is quiet most of the time. Teko shows up at moments that reward presence rather than filling silence. Every mascot interaction passes the doctrine before it ships.