Full depth — file names, mechanisms, and trade-offs.Concise and competent — what it does, without the internals.Plain language — the value of the work, no jargon.
Tools
Small self-contained things built in the lab. Free to use, nothing to install.Small free tools I've built. Nothing to sign up for, nothing to install — just open one and use it.
Voice Loop
A bring-your-own-key voice prototype that closes a full spoken loop around Claude using only the browser — SpeechRecognition in, speechSynthesis out, no server and no second vendor — and instruments the cascade: speech-end → transcript, transcript → first token, first token → first audio, and the total stop-talking-to-start-speaking, live per turn with a session median. Push-to-talk and continuous turn-taking, barge-in, and the raw transcript shown alongside every turn. One file, no build step.A bring-your-own-key voice prototype: talk to Claude entirely in the browser, using the browser's own speech recognition and speech synthesis. The point is the stopwatch — every turn is broken into the four stages of the cascade, live, with a session median, so you can feel where a voice agent's delay actually comes from. Push-to-talk or continuous turn-taking, barge-in, and the raw transcript shown next to the conversation so mishearings stay visible.Talk to Claude out loud in your browser and it talks back. The real point is the timer: it shows you exactly where the delay comes from — hearing you, thinking, then starting to speak — so you can feel why talking to an AI lags. You can hold a button to talk or just leave the mic on, and it always shows you what it thought you said, so you can spot when it mishears. You paste your own Anthropic key and it stays in your browser — though note the browser's speech recognition may send your audio to the browser maker.
Memory Inspector
A bring-your-own-key chat whose memory store sits open beside it. After each turn Claude runs an extraction pass that pulls durable facts about you into an editable localStorage memory list; the stored memory is injected back into the model's context every turn. Each entry's background shading is a 2-D read — confidence on one axis, sensitivity on the other — so what the system knows, and how sure and how private, is legible without opening a panel. Edit or forget entries inline; changes take effect next turn. Key stays in the browser. One file, no build step.Chat with an assistant that shows its memory next to it. After each reply, a Claude extraction pass adds or updates the durable facts it inferred about you, each shaded by inferred confidence and sensitivity. Edit or forget any of them inline; changes take effect on the next turn. Bring your own Anthropic key — it stays in your browser.You chat with an assistant, and everything it decides to remember about you appears in a list right next to the conversation. Each remembered fact is tinted by how sure the assistant is and how private the fact is — faint means unsure, red means personal — so you can see at a glance what it thinks it knows. You can fix a fact or delete it, and the next reply uses the corrected list. You paste your own Anthropic key and it never leaves your browser.
Branching Canvas Conversation
A bring-your-own-key chat laid out spatially on a canvas: every message is a draggable node in a conversation tree, and you can branch off any message into a parallel thread. Sending from a node uses the path from the root to that node as its context. Toggle between the canvas and a linear chat view. Vanilla pan/zoom canvas, Anthropic-only, key stays in the browser. One file, no build step.A chat drawn as a tree on a pan/zoom canvas. Each message is a node; branch off any message into a parallel thread, and the reply uses the root-to-node path as its context. Toggle a linear view. Bring your own Anthropic key — it stays in your browser.Normally a chat is one straight line. Here it's a tree you can see: every message is a card on a canvas, and you can split off from any point to try a different direction without losing the first one. You can also flip to a normal top-to-bottom view. You paste your own Anthropic key and it never leaves your browser.
Reasoning-trace A/B
A self-contained study on whether you can tell a sound explanation from a confident wrong one. An AI gives a question, a confident answer, and its reasoning trace; some traces are sound and some reach the same answer through a subtle fabricated or broken step. You judge which to trust, can point at the step you suspect, and find out how many flaws you caught versus how many fooled you — with a reveal-timing toggle (instant / delayed / on-demand). One file, offline, no model call.A study where an AI shows a question, a confident answer, and its reasoning trace. Some traces are sound; some reach the same answer through a subtle flaw — a fabricated fact, an invented premise, a broken justification. You judge which to trust, and see how many flaws you caught versus how many fooled you. A toggle varies when the trace appears (instant / delayed / on-demand). Runs entirely offline; the traces are hand-written and fixed.A little quiz: a computer answers a question and shows its work. Sometimes the work is honest; sometimes it gets the right answer using a step that's quietly wrong. You decide whether to trust it, then find out if you were fooled. It shows the point researchers found — a confident-looking explanation earns trust whether or not it's actually correct.
Generative UI Canvas
A bring-your-own-key canvas where a typed request makes Claude generate interactive HTML app windows, and clicking around inside a generated app posts back and re-invokes the model. Vanilla pan/zoom canvas, sandboxed iframes, key stays in the browser. One file, no build step, offline shell.Type a request and Claude builds interactive app windows on a canvas; interacting with a generated app posts back and re-invokes the model. Bring your own Anthropic key — it stays in your browser, and generated code runs sandboxed.You type what you want — a dice roller, a tip calculator — and Claude builds a little working app for it on a canvas. Clicking inside the app talks back to Claude, which updates it. You paste your own Anthropic key and it never leaves your browser; the apps Claude writes run in a locked box that can't touch it.
Cook Mode
A single-file recipe reader for cooking from a screen: one step at a time, timers parsed out of the step text, ingredient scaling and US/metric conversion, and a screen wake lock so the phone never sleeps mid-step.A single-file recipe reader for cooking from a screen: one step at a time, timers pulled out of the step text, ingredient scaling, US/metric conversion, and a screen wake lock.A recipe page made for actually cooking: it shows one step at a time, starts timers for you, doubles or halves the ingredients, converts between cups and grams, and keeps the screen from going dark when your hands are messy.