A curated catalog of real components with behavioral metadata, used to propose hardware component stacks. Full listing below.
⌘/Ctrl+Enter to propose a stack
Every listed object is real-world dimensions. Nothing is invented.
No stack yet.
Describe what you want something to do, set a constraint, or let ingenue surprise you. The parts assemble here. Five voices arrive on the right.
ingenue is a browser tab that treats a parts catalog like a vocabulary. You describe what you want something to do — in plain, weird, inexact language — and the engine proposes a set of real parts that together could be a real thing. Then five voices argue about whether it should be.
I built it because the gap between "I have parts" and "I have an original thing" is not an engineering problem. It is a design problem. Schematic editors assume you already know what you want. Spec-sheet search assumes you already know what to ask. Neither of those assumptions matches the way I start.
The name is the whole thing — an ingenue doesn't know the rules, that's the point, and by the time she learns them she has already made the combination nobody was going to try. The ingenuity root sits right there in the word. The Konami code is a nod to a dormant piece of gene-network software called Ingenue, from UW, early 2000s. I went to UW. Fine.
Add an API key in Prefs — your own Anthropic key, lives in your browser, I never see it. Without a key, the voices still speak from authored response banks. With a key, they get smarter and more specific.
Then pick a mode. Analogy takes plain-language descriptions ("something that knows when you've been gone too long"). Constraint is the Fuller move — set the envelope, see what fits. Curiosity is no input at all, and it's where ingenue earns its name.
It will not draw a schematic. It will not chase the lowest price. It will not tell you what's good — that's your job. The WiringMap is prose. The SkillPath names what to learn and stops. The point is to leave the session with something worth building, not with a ready-to-fab file.
ingenue borrows its visual paradigm from the Teachable Machine tradition — three modular blocks you can point at: what you want, what emerges, what is said. The typography is a reference-book register (Fraunces for display, Spline Sans for body, JetBrains Mono where things are technical). The NoveltyIndex readout is Kablammo — a variable font whose letterforms morph with the score. Typography as the chart. The whole thing is cream and viridian and, at the top end of novelty, pink — because surprise in this tool is cheerful, not incandescent.
Concept, domain expertise, UX authority: Phil Renato / renato.design
Code, algorithms, documentation: Claude / Anthropic
Methodology: ILCA — Iterative LLM Co-Authorship
Typefaces: Fraunces, Spline Sans, JetBrains Mono, Kablammo — all open-source, via Google Fonts
Catalog source: Adafruit Industries is the default (106 parts; unaffiliated; used as a corpus under fair-use observational analysis). Extended via Amazon (16), SparkFun (7), Pololu (4), Parts Express (2), PJRC (2), 80/20 (2), Digi-Key (1), McMaster-Carr (1), Sweetwater (1). 142 parts total — still a sliver of any one supplier's catalog, but enough vocabulary now for music players, toy robots, wearable jewelry, time-based kinetic sculpture, and a Pi or Jetson to host an LLM. The point remains behavior over coverage.
Early effort. ingenue is a young research artifact. Every number in the catalog — price, dimension, novelty weight — is an editorial claim, as close to right as I could get it. Supplier URLs drift; non-Adafruit links route through search results rather than direct product pages. Verify SKU, price, and availability before ordering anything.
Every part carries a real bounding box in millimeters. Click Envelope on the Engine module to see the stack as a physical object — packed top-down, sized against common things (pack of gum, Altoids tin, deck of cards). The reverse mode is Shrink-To-Fit — pick a target envelope and ingenue rethinks the whole stack inside it, with a Trade Report of what survived and what didn't. Every substitute is a real catalog part at its real size. Nothing is invented.
Any of the three modules — Input, Engine, FabricationVoices — can be double-clicked on its header to fill the workspace. A breadcrumb at the top shows where you are. Escape or click the breadcrumb to return.
The whole thing works on a phone, but a few desktop moves don't translate:
A parts catalog on a phone is imperfect. A laptop is better for this kind of thinking. But you can start the idea in a waiting room, then refine it at the desk.
Pick your LLM source. Without one, ingenue falls back to authored voice banks — still usable, just smaller.
Your key, your cost, your rate limit. Lives in this browser's localStorage only. Get one at console.anthropic.com.
ConstraintMode and CuriosityMode will only surface stacks at or above this score.
Reduced motion is respected automatically via system preference.
Clears API key, active voices, novelty floor, saved stacks.
Heads up: ingenue is brand new. If it reads you oddly, that's me still figuring out what I'm good at. Say it differently, or tell me more.
ingenue is a pipeline with five stops. You type a description, a constraint, or nothing at all. The IngenuityEngine reads your input against the PartsLibrary — a curated catalog of real Adafruit components with behavioral metadata. It proposes a ComponentStack — three to six real parts that together could be a real thing. The PackingEnvelope lays them out top, side, and in 3D, and tells you how big the thing would actually be. Five FabricationVoices — The Tinkerer, The Professor, The Artist, The Engineer, The Bricoleur — argue about whether it should exist. You leave with an ArtifactSpec: concept, parts, wiring in prose, skill path, voice notes, open questions.
The LLM is pluggable. Anthropic Cloud by default; Ollama or LM Studio if you want it local. Without an LLM at all, the voices still speak — from authored fallback banks. Fewer lines, same shape.
With the current catalog (36 parts, growing), ingenue can propose systems that:
(capabilities summary renders here at runtime)
Every stack exports a full report (.txt) that includes a step-by-step assembly guide: what tools you need, what order to solder/wire in, how to verify each part in isolation, and a first-power-on checklist. The Adafruit BOM (.csv) lists every SKU so you can build the cart manually.
Honestly named limits, so you know where the tool stops:
When you ask it to peel a potato, it will tell you it can't peel anything.
It might offer to detect when you've peeled one, or to time how long it took.
That is the point.