Each academy is three projects. Difficulty climbs. By the third project of any academy, you've shipped something a real person can use.
The first academy. Three projects, each one converts something that lives only in your head — a rabbit hole, a hobby, a weird competence — into a Claude Skill that actually works for the person you built it for. By project three, you ship a versioned, tested, refusing-when-it-should Skill that other people install and use.
Pick a domain you know more about than the adults around you. Build a Claude Skill that thinks the way you do about it — with your taxonomy, your edge cases, your voice.
Three Skills that call each other and share memory, designed for one specific person you can name. They have to feel like one thing, made for one person.
Write tests for your Skill. Evolve it without regressing. Teach it when to refuse. Ship a v1.0 a real user installs.
Real HTML, CSS, JavaScript, real APIs, real tests. No drag-and-drop. No toy syntax. By project three, you have a tool deployed on a real URL that survives a week of being used by the person you built it for. You're allowed to use Claude the entire time — the rule is only that the part that's only you must be visibly only you.
A page on the internet, built with HTML/CSS/JS, that does something useful for one specific person in your life. Deployed to a public URL by the end.
Add real intelligence: parse JSON, call the Claude API from your own code, handle errors, stream responses, keep keys safe.
Tests, edge cases, error states, accessibility. Ship a tool, hand it to the person you built it for, and survive a week of real use.
The most empathetic academy. Agents take action on someone's behalf — and that "someone" matters. You'll learn to design from observation rather than assumption, give your agent permission tiers, build a disagreement protocol, and finally — in project three — build an agent for someone genuinely vulnerable. The point of this academy is that the right answer is sometimes "don't."
Pick someone real. Watch them for a week — what do they actually do? Build an agent for that, with an audit trail every action is reviewable on.
Permission tiers, scope, escalation. Build a 5-level trust gradient and a disagreement protocol. The agent has to know when to refuse you.
The capstone. Build a real agent for someone genuinely vulnerable — your grandmother, a sibling, a friend going through a hard time. Ship something they actually use, with a quiet-harms checklist passed.
The deepest academy. Where the other three teach building things, Harness Studio teaches building systems that keep other systems honest — including your own drifting taste. You define what "good" means, build a judge that scores work against your rubric, and finally, in the grand finale, you weave together a Skill, a tool, an agent, and a judge into one system that catches you drifting from your own standards over time.
Pick a domain you care about. Write the rubric — five dimensions, with anchors. Score five real examples by hand. Notice what your taste actually is.
LLM-as-judge, calibrated against your hand-scored examples. The judge has to agree with you 80% of the time before it ships.
Run your judge against your own work, weekly, for a month. See yourself drift. Add the Skill, the tool, the agent. The grand finale weaves all four academies into one system.