About PracticeTypes.ai
Clinicians have the best ideas in healthcare. Almost none of them ship.
I've spent the last decade building technology alongside clinicians — sitting between the exam room and the product team, watching the same failure repeat: the person who understands the problem can't show what they mean, and the people who can build never quite hear it. Ideas die in translation. PracticeTypes.ai is my attempt to close that gap at scale: teach licensed clinicians to prototype their own ideas with AI, so they walk into the next conversation with a working thing instead of a description.
The product is simple to experience — describe a tool you wish existed, watch a clickable prototype build itself in minutes, then learn to take control of it — but the choices underneath it are deliberate. This page is where I show the work.
What we believe
The prototype is a communication tool, not a product.
A clinician's prototype doesn't need to ship to an app store. It needs to make an IT team, a CMO, or a co-founder finally see the idea. That framing changes everything downstream — including sidestepping PHI entirely: we build with realistic synthetic data, never patient data, and remind users of that at every upload.
The skill is the promise; the artifact is the hook.
Prompting an AI collaborator well is a career skill now. But nobody learns from a lecture — so the first fifteen minutes deliver a working prototype of your idea, and the course teaches you to take ownership of it.
Function before looks.
First drafts put the effort into the workflow — real buttons, real flows — dressed in a starter palette Kit picks to fit your world. Applying your own brand is a deliberate, earned moment that comes next.
Community compounds.
Every prototype a clinician shares with colleagues is proof that people like them can build. The long game is a community of clinician-builders sharpening each other's ideas — the part of this that gets more valuable every year.
How it's built
PracticeTypes.ai is an agentic product, and the architecture follows a rule we learned the expensive way and the field has since converged on: deterministic workflow in code, bounded AI judgment inside it.
The flow never asks the model what comes next.
Every stage transition — build, questions, enrollment, path choice, style pass — is owned by application code and triggered by user actions. Kit, our AI guide, reacts and advises; she never advances state. This is the difference between an agent that feels reliable and one that drifts.
Kit is one personality, many specialized brains.
An enduring voice layer defines who Kit is everywhere; each segment of the workflow — clarifying questions, conversational triage, live patch edits, full generation — gets its own purpose-built prompt on top. Her triage brain routes every message to talk, edit, or rebuild, so a frustrated or vague message gets a colleague's reply instead of a tool error.
Memory with a writer split.
The project's durable memory is written by deterministic code for every mechanical fact — what was built, what was decided, what was paid for — while the AI contributes only bounded judgment fields through a single parsed channel. Facts can't drift; judgment stays useful.
Healthcare posture by design.
Synthetic data that would make a clinician nod — right vocabulary, right units, right workflow states — with a no-PHI courtesy at every upload surface. We teach expressing ideas safely, not shipping clinical software.
Build journal
Running notes as we build — decisions, features, and the lessons that changed our minds. Newest first.
- decision
A tester rewrote our roadmap: color-first drafts and a curriculum you can't see
We've been gathering feedback from the people who actually went through our first prototype, and it's driving the roadmap. This week we walked one of those testers — a designer — from the old version to this one. Her verdict on the rebuild ('a 10x improvement') mattered less than her criticisms, because they echoed what other prototype users had hit and pointed at decisions we'd been proud of. So we reversed them. Grayscale-first is gone: first drafts now land fully designed in a starter palette Kit picks to fit your world, because the first render is the moment someone decides the idea is real — and 'function before looks' survives as what it always really was, effort spent on the workflow, with 'make these colors truly yours' preserved as its own moment later. The bigger reversal: the visible curriculum rail is gone. A checklist down the side made the course feel like a tug-of-war — pulled forward before you were done, or toward lessons that weren't what you cared about right then. The curriculum still exists, as a dependency graph running in the background; what you see is a choice: at each natural pause, Kit offers the two or three skills you've unlocked, plus two standing options — 'just keep building' and 'what should I add next?' That last one came straight from the tester's sharpest insight: some people's blocker is knowing how to prototype, but for many it's knowing WHAT to prototype — so Kit now proposes concrete features grounded in your world, not just edits to what exists. And because every fork is now a real choice, every fork is also a measurement: we record what was offered and what got picked. Enough builders through the path and the optimal sequence stops being our guess — the curriculum learns from the people using it.
- lesson
The AI couldn't see the website — so we stopped asking it to look
'Match my website' kept failing: wrong colors, invented logos. The root cause was invisible until we traced it — the AI's web-fetch tool returns a page's TEXT, not its code. No CSS, no hex values, no font declarations, no logo files. The model literally could not see the brand it was asked to copy, so it guessed. The fix inverted the architecture: our server now fetches the raw HTML and stylesheets itself, deterministically mines the evidence (color frequencies, font-family declarations, the actual logo file URL), and the model only ranks what was observed — inventing a color it was never shown is now structurally impossible. First test against a real practice site returned the exact brand palette, the real display font, and the genuine logo SVG.
- milestone
Every lesson is real now
All twenty lessons shipped with full bodies — reformulated from our previous prototype's education library into interactive modules in Kit's feed: quick-check drills that bank themselves when you answer right, one-click prompt starters that load into the chat composer for you to edit and run, and worked examples of what good looks like. The placeholder era ('the full version is coming') is over.
- decision
Curriculum v2: the course became reps, not reading
Rebuilt the curriculum from our previous prototype's education library, reshaped for a product where the lesson and the build share one screen. Every skill is now a rep on the learner's own prototype: edit a best-practice-annotated prompt Kit drafts for you, dictate a change, drop a napkin sketch, file a real bug report. A numbered best-practices thread runs through the path, and the tooling gauntlet (GitHub, deployment, terminal agents) moved out of the course body entirely — to an offramp fork where owning your code is an opt-in graduation at peak motivation, not a mid-course toll. Everyone leaves with a downloadable skills reference, paid or not.
- decision
Memory architecture: code owns facts, the AI owns judgment
Researched current best practice for long-horizon agents (structured note-taking, compaction, provider-native memory tools) and locked our memory design: a project file that is JSON-canonical, rendered to markdown, injected into every one of Kit's brains. The rule that matters: deterministic code writes every mechanical fact (zero drift), and Kit writes only bounded judgment fields through one parsed channel. We deliberately did NOT adopt the platform's agent-managed memory tool — handing the model write authority over facts reopens the exact drift class the split closes.
- lesson
A paying member got treated like a free user — and why
Kit refused a paid member's edit, citing the free-tweak limit — even though the enrolled flag was correctly in her context. The model trusted the conversation history (where pre-paywall Kit had legitimately cited limits) over the state flags. The fix: enrolled state now explicitly declares earlier limit-citations obsolete. The general lesson for agentic products: when state and history disagree, models follow history unless you tell them not to — loudly.
- feature
The curriculum moved into the conversation
Lessons now render as interactive modules inside Kit's chat feed — buttons, choices, and forms arriving in the conversation flow — instead of a separate content pane competing for attention. One conversation surface, one prototype surface, one navigation rail. The chat anchors to the newest message even as panes resize around it.
- feature
Kit shows her work
When Kit edits the prototype, every changed element gets a shaded highlight with her K badge, all at once, fading slowly — and the canvas scrolls to the first change. Trust in an AI teammate is built in exactly these seconds: the user asked for something, and the interface points at the proof.
- lesson
Sketch-style wireframes made people afraid to click
Our first drafts rendered in a deliberate Balsamiq-style sketch aesthetic — hand-drawn fonts, wobbly borders — to signal 'this is early, judge the workflow.' Dogfooding revealed the cost: sketches read as pictures, and people don't click pictures. The reversal: drafts stay strictly grayscale (function before looks survives) but with real buttons, real links, hover states. Affordances beat metaphor.
- feature
The whole journey became one conversation
The landing page's idea form is now literally Kit's first chat message; the visitor's idea is their reply; and when the workspace opens, that exchange is already at the top of the thread. No cold-open agent, no context reset — the person who asked you about your idea is the one building it.
- decision
Adopted the AI-standard shell
Replaced our novel picture-in-picture layout with the shape every AI tool has trained users on: navigation rail left, conversation center, live artifact right. The clever layout was costing cognition that the familiar one gives away for free. Enrollment became pure addition — the rail and lessons arrive around a prototype that never moves.
- decision
Wireframe-first drafts
First drafts render in grayscale on purpose — the discipline professional product teams use to separate how it works from how it looks. It also absorbs the aesthetic-dissonance problem (an AI guessing your brand wrong erodes trust) and makes 'your style, your logo, live' the first thing the paid course delivers.
- milestone
PracticeTypes.ai is live
Domain to deployed product in two days: idea in, working grayscale prototype out in minutes, with Kit asking sharp questions in parallel and folding the answers into the build. The stack is deliberately lean — Next.js, one model provider, server-side web tools, static-file prototypes — because speed of iteration is the whole game at this stage.
- milestone
The pivot: clinicians
Cohort 1 of the general-audience version taught us the thesis survived but the vehicle failed: interest was real, yet zero of twelve finished — first gratification came too late, and a chat-only interface made people do the work of driving. PracticeTypes.ai is the rebuild around those lessons, for the audience where a prototype has a natural payoff: licensed clinicians with real problems and a specific colleague to show.
The best way to understand it is to use it.
Try it free — describe a tool you wish existed