Foundation

Curricula are downstream of kids who love to build. We align to standards because it opens the school motion, not because it defines the product. Every time we have softened a kid experience to match a district-procurement checklist, the kid experience has lost and so has the procurement cycle. Kids first, in sequence and in weight.

Concrete example. The Stage 1 regex concept detector lags a real LLM classifier on coverage. We ship the regex anyway because the weekly parent email is more important than the perfect tag graph. The LLM classifier is TODO-002 for Stage 2. Kids are using the tool now. The curriculum alignment deepens on a schedule that does not interrupt them.

TODO(foundation): Lawless to confirm wording and ordering.

Parents do not buy kid software based on a landing page. They buy it based on the first time the AI handled a flagged prompt correctly, the first time the weekly email showed a concept their kid actually learned, the first time a mistake was recovered without shame. The bank account builds one transaction at a time. We take the account seriously.

Concrete example. The weekly parent email does not go out until a human has reviewed the first 50 for tone and accuracy. The automated send lives behind a feature flag until the parent NPS gate clears. We are willing to be slower to ship than a growth playbook would suggest because the cost of the wrong first email is the account.

TODO(foundation): Lawless to confirm wording and ordering.

The difference between Tekku and every coding app in the category is the URL. A kid with a deployed app has proof. The kid can text it to a grandparent. The grandparent can use it. The feedback loop is human. Everything we build is in service of getting the kid to the URL faster and more often.

Concrete example. The Vercel Blob plus Next dynamic route ship pipeline is the fastest path to a hosted, shareable artifact that we can own end to end. We rejected fancier infra plans in Stage 1 because the goal is the URL on day one, not the infra that would host the URL at scale. Scale follows the habit.

TODO(foundation): Lawless to confirm wording and ordering.

A kid-AI product that grows faster than it tunes its moderation ends up on the wrong news segment. We would rather hit a smaller pilot cohort at the right quality than a larger one with a predictable incident in month four. Scale is the second-order metric. Safety posture is the first-order one.

Concrete example. The Stage 2 COPPA stack is a hard gate for expansion beyond founding families. The moderation false-positive tuning layer is TODO-003 in the backlog, blocked on four weeks of real-data tuning. We will not flip on open signup until both are in place and we have run the transparency page past counsel. Growth waits.

See also the safety model page for the full posture.

Every AI-in-education story right now is a story about teachers losing. Tekku makes teachers win. The Workshop tier gives a teacher classroom-ready projects that align to standards without asking them to become a prompt engineer. The same teacher then goes home and can explain to her own kid what the tool does. Ally, not opposition.

Concrete example. The Stage 2 pilot design is 20 to 50 kids across three classrooms, no paid conversion, full access to the weekly kid reports and aggregate classroom view. We are not selling. We are listening. The procurement cycle comes after the teacher has told two other teachers unprompted.

TODO(foundation): Lawless to confirm wording and ordering.

Every Tekku sales surface, pitch deck, and investor conversation runs through the live product. The founder opens Tekku on the call, gives the prospect an 8-year-old name, and builds something. The demo is not a reel. It is the actual tool. This constraint forces the product to always be demo-ready, which forces it to always be kid-ready.

Concrete example. This investor portal renders live numbers from the model sheet every time it loads. The hero metric strip is not a static image. The scenario badge in the header shows which scenario the sheet is currently serving. If the numbers move, this page moves with them in the same render cycle. The principle shows up in the code, not just in the voice.

TODO(foundation): Lawless to confirm wording and ordering.

The compounding unit at Tekku is the family, not the user. A family that is happy pulls another family in through a shipped app, a school conversation, a grandparent who wants the same thing for another grandchild. That is the distribution we build for. Paid acquisition exists to find the first families, not to replace the loop once it is running.

Concrete example. The viral badge on every shipped app is sized for readability on a phone, links back to tekku.app with a referral slug, and is deliberately designed to be embedded in an app that looks nothing like Tekku. We want a grandparent to play a clap counter on an iPhone and click the badge. One grandparent at a time.

TODO(foundation): Lawless to confirm wording and ordering.

The first four human hires are named by the founder, not by a recruiter. Every hire has a specific trigger, not a general contribution. Founding eng 1 owns the workspace state machine. Founding eng 2 owns the AI invocation layer. Head of trust owns the COPPA posture end to end. Head of school GTM opens the Workshop tier at the Stage 3 gate. Named ownership first, headcount later.

We hire for people who have shipped inside a kid-facing surface or a safety-critical surface. Game studio, education tech, trust and safety, moderation tooling. The pattern is a builder who has felt the weight of a 12-year-old using the thing they made.

Compensation is cash plus equity at the market line for a seed-stage company with a credible raise. No tricks on vesting. Four-year vest, one-year cliff, quarterly after that. Remote-first with one in-person week per quarter at a shared location. The document that describes this lives in the data room and is given in full before an offer is signed, not after. TODO(foundation): Lawless to confirm remote-first posture and quarterly cadence.

To parents, we write at grade 8 to 10. We say what the product did, what the kid did, and what the parent can do next. We do not use the word "learning" as a verb for things that did not happen. We avoid adjectives where a number will do. The weekly email is the test: if a parent forwards it to her own mother, we got the register right.

To kids in-product, we write at grade 5 or below. Errors are translated into actions. "Your code got stuck. Let's go back to when it worked." Checkpoints are binary. "Do you want both players to share one button, or should each player have their own?" Red is banned. The punch-warm palette handles warnings. Kid copy ships after a kid reads it out loud.

To press, we speak on the record, with the product open. Every press conversation includes a live demo under the name of a hypothetical kid. No photoshoot quotes. To regulators, we speak in writing, with counsel in the thread. We do not pre-empt. We respond. The transparency page linked from the safety model is the public anchor.

We collect only what the product needs. A kid profile carries a nickname, an age, and a parent linkage. Transcripts live for 90 days, then the retention cron deletes them. Snapshots follow the same window. Project files persist until a parent deletes the account, at which point everything cascades out. Parent-facing export and delete controls ship with the Stage 2 COPPA stack.

We do not train on kid work. The clause is written into the Anthropic and OpenAI contracts we use at the API tier, and it is the first sentence of the privacy policy. Moderation events are stored for policy review, not for model improvement. When we eventually train a concept classifier, it runs on synthetic data and on anonymized aggregate labels, never on raw kid transcripts.

The full posture, including the retention cron schema, the parent verification path, the moderation review queue, and the public transparency page, sits on the safety model page. This section is the summary. The safety model is the contract.