
conceptcraft
A Claude Code skill for scroll-driven zoom-canvas explainers: one concept, one engraved drawing, a camera that zooms, a finale that runs the system live unde…
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up conceptcraft (claude-skill project) into my current project. Found on https://claudeers.com/conceptcraft Repo: https://github.com/maraja/conceptcraft Homepage/docs: — Detected install method: claude-skill → # copy this skill into .claude/skills/conceptcraft/ Category: devtools. Platforms: cli, api, desktop, mobile. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: active; community-verified: false. Confirm the source before running anything.
# copy the skill dir into your project: # .claude/skills/conceptcraft/ (or ~/.claude/skills/conceptcraft/ for all projects)
git clone https://github.com/maraja/conceptcraft
// compatibility
| Platforms | cli, api, desktop, mobile |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | NOASSERTION |
| Pricing | open-source |
| Language | JavaScript |
conceptcraft
A Claude Code skill that builds scroll-driven zoom-canvas explainers: one concept, one engraved drawing, a camera that zooms. Scroll drives everything — the camera walks the concept stop by stop, each stop expands from a quiet mark into a labeled diagram with a readable artifact panel, a live HUD instrument tracks the concept's own state, and the finale runs the whole system deterministically under scroll, including one caught failure.
The output is a single self-contained artifact.html you can host anywhere:
a static directory on your site, a Claude artifact, an S3 bucket. No build
step, no dependencies at runtime, fonts from Google Fonts, everything else
inlined. Built for teaching: classroom lectures, conference talks, YouTube
companions.
Install
git clone https://github.com/maraja/conceptcraft ~/.claude/skills/conceptcraft
That's it — Claude Code picks it up as a skill. Then, in any session:
make a conceptcraft explaining how RAG retrieval works
Claude interviews you briefly (the stops, the depth, the HUD instrument, the finale), scaffolds a build from the reference implementation, rewrites the authored layers for your concept, and runs a screenshot-based verification harness (desktop at two heights, mobile, reduced motion, contrast) before shipping.
What's in the box
SKILL.md the format laws + workflow Claude follows
references/MECHANICS.md every subsystem, split MECHANICS vs AUTHORED
references/MODES.md the four presentation modes + how to pick one
references/modes/ copy-paste panel kits, one per mode
references/SHEETS.md the five sheet worlds + the furniture switch
references/sheets/ one token block per world, and the switch
references/DEPTH.md the three planes + the veil that makes them
references/depth/ depth.css + depth.js, the compositor kit
references/GOTCHAS.md the debugging table (every row cost real time)
references/reference-build/ two full reference pages (concept + worked example)
engine/ the scrollcraft runtime (vendored, never edited)
scripts/scaffold.mjs one command starts a new build
scripts/serve.mjs tiny static server for verification
scripts/shoot.mjs the screenshot/contact-sheet verification harness
scripts/probe.mjs screenshot + live state at exact scroll positions
scripts/svgcontrast.mjs on-screen contrast for SVG sheet labels (+ px.mjs)
The only tooling dependency is playwright-core (installed per-build, used
only by the verification harness).
The format in one paragraph
At rest, every stop on the sheet is an icon — a muted mark and a dim name. Only the scroll-current stop expands into a full diagram, driven by focus trapezoids of scroll progress. Labels are screen-sized and haloed, placed by hand per stop. The drawing narrates a real example (real filenames, real numbers), never generic part names. After the guided tour, the page runs the system it just taught — deterministically, as a pure function of scroll, with accumulating state and one caught failure — and what the run resolves stays lit through the close. Every frame of that is verified from screenshots before it ships.
Sheet worlds: how much of the page is furniture
The reference sheet is a dark engraved plate carrying five furniture layers
at once: a fine grid, a coarse grid, a frame, crop marks and a title block.
None of them carries information, and inheriting all five everywhere reads
as busy. Furniture is opt-in: references/sheets/_furniture.js replaces
the hardcoded block with a SHEET config, and each world ships a token block.
| World | Ground | Furniture | Best for |
|---|---|---|---|
| Blueprint | dark olive | all five layers | subjects that genuinely ARE instruments |
| Clean paper | light warm | none | calm; video and print |
| Chalkboard | matte slate | none, chalk strokes | teaching in a room |
| Editorial | warm paper | two rules at most | pages that read as writing |
| Bare stage | near-black | none | one strong drawing |
The three axes compose: a page is a theme, a mode and a world. Going light
inverts more than colour, and a transparent card changes the composition,
so read SHEETS.md before switching an existing build.
Modes: the vessel matches the material
A conceptcraft has three independent axes, plus depth on every build. Theme is colour temperature, so a series never looks identical. Sheet world is the ground and how much furniture it carries. Mode is the information-display language, and it is chosen from the TYPE of material being taught:
| Mode | Material | Panels look like |
|---|---|---|
| Terminal | code, commands, logs | dark well, mono lines, syntax colour |
| Archive | history, people, documents | typeset excerpts, citations, highlighter |
| Notebook | math, derivations | ruled paper, worked steps, margin notes |
| Field Guide | comparisons, data | numbered plates, specimen boxes, charts |
One dominant mode per page, declared as mode--<name> on the act root; a
single card may borrow another mode's panel when its material calls for it.
The engraved sheet, camera, and instrument never change. This exists because
a history page once shipped a 1973 government report inside a syntax-coloured
terminal well: a quotation is a document, and it should look like one.
Depth: three planes, never one
Theme, mode and sheet world decide what the page is made of. Depth decides how far away each thing is, and every build gets it.
Every card-up frame carries three planes: background (the furniture and every resting stop, washed toward the ground colour and thrown out of focus by a backdrop blur), midground (an aperture cut over the current subject, left untouched), and foreground (the card and the chrome, sharp, on a field that has gone quiet). One element does all three.
This replaced opacity, which cannot separate planes: a label at 0.38 is still
a sharp label at full size competing for the same eye. It also retires the
radial vignette the format used to carry, most of the focus plate, and the
reason builds needed hand-measured camera-pan tables to keep the drawing off
the copy. references/DEPTH.md has the five rules and the per-world wash
tokens; the two traps that cost the most are mounting the veil on <body>
(the pinned stage is a stacking context, so the veil then blurs the cards it
was meant to lift) and ramping the veil's opacity (a backdrop-filter
element at opacity < 1 composites its blurred copy over the sharp original,
and the sheet arrives doubled).
Topologies
The reference page draws a ring (an agent loop), but the format is topology-first: a pipeline is a path, a hierarchy is a tree, a shared-memory system is a hub with its operations cycling around it. The skill's first job on a new concept is decomposing it onto its TRUE shape.
Credits
- Format and reference builds by Amit Maraj.
- The scroll engine (
engine/scrollcraft.js+.css) and the verification harness (scripts/shoot.mjs,scripts/serve.mjs) are from scroll-craft by Nate Herk, MIT licensed, vendored unmodified.
License
MIT — see LICENSE.
// faq
What is conceptcraft?
A Claude Code skill for scroll-driven zoom-canvas explainers: one concept, one engraved drawing, a camera that zooms, a finale that runs the system live under scroll.. It is open-source on GitHub.
Is conceptcraft free to use?
conceptcraft is open-source under the NOASSERTION license, so it is free to use.
What category does conceptcraft belong to?
conceptcraft is listed under devtools in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/conceptcraft)
// retro hit counter
[](https://claudeers.com/conceptcraft)
// reviews
// guestbook
// related in Developer Tools
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Curs…
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA
AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs,…