Sp Canvas

Prototype canvas

A local tldraw app that discovers every .html file under canvases/<slug>/ and renders it as a shape. There is no shape map to edit and no code change needed to add a board.

A project is a folder under ~/Documents/Super Prototyping (PROTOTYPING_PROJECTS_DIR moves it), and its boards are in that folder's canvases. No folder anywhere else is ever a project, so deleting one from the app can only ever trash what was made there. Working in another repository, write the boards to ~/Documents/Super Prototyping/<project>/canvases/<slug>/, not to the repo. The app is installed outside every project, so upgrading it never touches a board you wrote. Every launch of the app puts sp, refkit and artgen on PATH and links these skills into each agent's skills directory.

Start

bashCopy code
sp open

That opens the Super Prototyping app on its home page, starting it if it is not running, and prints the canvas address and the projects folder. It is the way to show the user the canvas. A project's boards are at <address>p/<project>/.

sp not found? Run sh <this skill's dir>/scripts/install.sh (Windows: powershell -ExecutionPolicy Bypass -File <this skill's dir>/scripts/install.ps1). It installs uv and the app and opens the app, which links sp. Then follow references/install.md to put it on this shell's PATH and go back to the task.

[super-prototyping:notice] on stderr from sp, refkit or artgen means the app found a newer release. The line says what to do: finish the step you are on, run sp upgrade, and tell the user what it printed. Do not upgrade in the middle of a step, and do not stop the user's task over it.

sp start serves the canvas without the app, for where it cannot run: CI, Linux, a remote box. On first run it fetches the canvas built for its version into ~/.cache/super-prototyping/<version>/, serves it on 127.0.0.1:5173 with node or bun, waits for the port to bind, and prints the address. A checkout being worked on serves its own canvas/dist instead, rebuilt with bun when a source is newer. It refuses a port that already answers rather than reusing it, and says so when that port is the app's.

• Boards are canvases under a project in the projects folder, each project served at /p/<name>/. The address printed is the home page.
• Port with --port N, or SP_CANVAS_PORT for a machine that always uses another one. A port that already answers is never reused: it may be another project's canvas, so start refuses rather than showing you the wrong boards.
• Two sp starts on two ports can run two canvases. The session name, the log and the pidfile are all keyed by port, so a second start on a free port leaves the first one alone. stop and status take --port for the same reason, and stop only ever kills the canvas it started.
• sp root prints which tree it found, the app's or a checkout's, and with -v everywhere it looked. The first thing to run when the canvas is not what you expected. sp status says whether the app is running, and on which port.
• It writes two directories: that cache, and ~/.local/state/super-prototyping/ for its pidfile and log, the same on macOS as on Linux; the app keeps its own state beside them. SUPER_PROTOTYPING_HOME moves both under one root. sp paths prints them and every variable in use; sp clean removes the cache and start's own files, never the app's.
• Deep-link a page with ?canvas=<slug>, e.g. http://127.0.0.1:5173/?canvas=notion-ios, and one board of it with #<file> after that, e.g. ?canvas=notion-ios#02-search-ask-ai: it opens selected, zoomed to fill the window. Give the board link when pointing at one screen.
• Read a board as a web page. The "Export to Figma" button in the top bar opens every board of the page in one scrolling document, each at its own size, at sheet.html?canvas=<slug>, and one board is at /board/<slug>/<file>.html. Both are ordinary addresses serving the board's own HTML, so they can be linked, reloaded, and read by the browser extensions — a Figma importer, say — that refuse to work on a generated page. This is where to read type at the size it ships at, rather than at whatever the canvas is zoomed to. The sheet opens on what to do with it: the html.to.design browser extension, which captures a localhost page that the Figma plugin's own servers cannot reach, and the paste or plugin route from there into a file.

Keep it on loopback. This is a local design tool, not a service to expose.

A project with no boards yet opens on a notice naming the directory the canvas resolved, rather than an empty grid: an empty boards folder and a canvas pointed at the wrong one look identical otherwise.

A folder created after boot appears on its own. The server watches the boards directory and rebuilds its index when a board folder or file is added or removed. Rewriting a board swaps the new version in without reloading the page, so a generator can be re-run with the canvas open while the person keeps drawing on it. If a ?canvas=<slug> link still matches no page, the folder has no .html file in it yet — an empty folder is not a board.

The styles panel is hidden by default; toggle it from the toolbar. Always-snap is on by default. The setting is per browser, so turning it off in tldraw's preferences menu sticks.

Boards

One folder under the boards directory = one tldraw page; one .html file = one shape. Switch with the page menu at the top-left; do not build a separate switcher. references/layout.md has the layout.json schema, the caption rules, and the 478 × 980 / sandbox constraints every artboard lives under.

Write nothing at the project's root but its documents (*.md) and project.json: work goes in a canvas folder, and what a run makes in that folder's scratch/. A project is what gets shared, and sp pack packages the root's canvases/ and documents and nothing else there. Keep project.json's id and format as they are. The layout is references/layout.md, under "The project folder". Share one with the community only when asked, as it says under "Sharing a project".

After editing layout.json, right-click the canvas and choose Force refresh. Shape creation is idempotent. It fills in what is missing but never moves a shape that already exists, so inserting or reordering a row entry leaves the old shape at its old position, overlapping the new one. Force refresh deletes every canvas-file / canvas-row-heading / canvas-file-label shape on all pages and rebuilds them from the current files. Content-only edits to a placed file do not need it: the canvas swaps the rewritten board in by itself.

Place things on the canvas

sp canvas reads what is on a canvas and places shapes, images, video and boards anywhere on it. The project's canvas open in the app runs each command with tldraw's own editor and answers once the result is saved to canvas.json, so a command that returned is on disk.

bashCopy code
sp canvas <op> --canvas <slug> '<json>'     # - instead of the JSON reads it from stdin

The chat panel's agent already has the project and the port. From anywhere else, add --project <name>, the name in the canvas's address (/p/<name>/). A command needs the project open in the app. With none open it waits ten seconds for one, then fails and says so: ask the person to open the project (sp open, then the project) and run it again. Never write canvas.json yourself instead.

What create takes:

• Image or video: {"type": "image", "x", "y", "file": "<path on this machine>"}, or "video". sp copies the file into the canvas folder's files/ first. Give w or h to scale it evenly, both to set its size, neither to keep its own.
• Board: {"type": "canvas-file", "board": "<file>.html", "x", "y"}, a board of this canvas folder, at its own size unless w and h say otherwise. Place only a board that no layout.json row lists, or it shows twice. One you placed stays out of the grid of unlisted boards.
• Text, note, geo: {"type": "text", "x", "y", "text": "..."}. Other props go under props, never w/h at the top level.
• Arrow: {"type": "arrow", "x", "y", "from": "<id>", "to": "<id>"} binds both ends, so it follows them. It cannot bind to a layout shape.
• An id of your own ("shape:keyframes-title") lets a later command name the shape. Every command is checked whole first: one bad shape and nothing is written.

Work out coordinates from get, never guess them. To put something beside a shape, start at its right edge plus 40 and check the new bounds against every other shape's, moving down past any it would cover. The person's own content is never to be covered.

What you create is yours to move and to change, and the person's to move and resize too: images and video keep their proportions, a board stretches freely (Shift keeps them). Content layout.json places stays locked. When the person has moved a shape of yours, theirs is the last word: an update or a layout op that would move it again fails with moved_by_person, listing each shape's bounds now. Run get again and plan around where they put it. A shape that only went along with the frame the person dragged has not moved. "force": true overrides it; use it only when the person asked for the shape to be put back.

Refusals come back as JSON on stderr with a non-zero exit, {"error", "message", ...}: not_agents (not yours), moved_by_person, holds_persons_shapes (the person dragged a shape of theirs into your frame), locked (the person locked it), not_current_page (the person is looking at another canvas), exists, not_writable, bad_command, and failed (the page could not save it; say so rather than retrying blindly). Your changes stay out of the person's undo.

Six keyframes of the person's screen recording, from $WORK/frames, put in a frame beside it with an arrow from the recording:

bashCopy code
sp canvas get --canvas video-notes
# {"id": "shape:rec", "type": "video", "owner": "person", "x": 0, "y": 0, "w": 720, "h": 405, ...}
sp canvas create --canvas video-notes - <<EOF
{"shapes": [
  {"id": "shape:k1", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/1.png"},
  {"id": "shape:k2", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/2.png"},
  {"id": "shape:k3", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/3.png"},
  {"id": "shape:k4", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/4.png"},
  {"id": "shape:k5", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/5.png"},
  {"id": "shape:k6", "type": "image", "x": 760, "y": 0, "w": 240, "file": "$WORK/frames/6.png"}]}
EOF
K='"shape:k1", "shape:k2", "shape:k3", "shape:k4", "shape:k5", "shape:k6"'
sp canvas pack --canvas video-notes "{\"ids\": [$K], \"gap\": 16}"
sp canvas frame --canvas video-notes "{\"ids\": [$K], \"name\": \"Keyframes\"}"
# {"frame": "shape:…"}
sp canvas create --canvas video-notes '{"shapes": [{"type": "arrow", "x": 0, "y": 0,
  "from": "shape:rec", "to": "shape:…"}]}'
sp canvas shot --canvas video-notes -o scratch/canvas.png

Then get again and check the frame's bounds against every other shape's before telling the person it is done.

What the person put on the canvas

Anything that is not a board or a layout.json picture or card is the person's, what they pasted or dropped, or yours, what you placed with sp canvas (owner in get says which). It is saved in the canvas folder:

• canvas.json: tldraw records (shapes, bindings, assets), sorted by id. A shape at the page's root has no parentId.
• files/: the files those records point at, by asset id. An asset's src is relative to the canvas folder (./files/<name>).

Read them to see what the person put there. A chip or a link named <slug>/canvas.json#<shape-id> points at one record, and one named <slug>/files/<file> points at one file. Never edit either. To answer a sketch or a note, place a board or a shape beside it with sp canvas. What it creates is stamped meta.by: "agent", and every write refuses a shape without that stamp.

Never let sp canvas commands inject arbitrary JavaScript, never load untrusted HTML into a board, and never add allow-same-origin to the artboard iframe.

Annotated screenshots

The review loop is a screenshot of the canvas with notes drawn on it, pasted into chat. Boxes, arrows or numbers all work, from tldraw's own draw/text tools or any image annotator.

1. Treat each annotation as an exact visual target, and say back what you read it as ("box 2: tighten the card gap") before touching anything.
2. Read the surrounding UI and the HTML source before editing.
3. Make the smallest source change that satisfies it.
4. Let the canvas pick up the board, then verify the same region visually.

Do not build an annotation-to-agent protocol. The screenshot is the bridge.

The chat panel

The canvas has a panel on the left, in the app or under sp start: a message to Claude Code or Codex — the mark on the header picks — Claude with its permission prompts off and Codex in its workspace sandbox, and what it did as it happens. The panel names the project and the canvas in front and tells the agent, and points the agent at this skill before it touches a board folder. A board it rewrites appears in place as any rewrite does, without reloading the page.

A conversation is a session the agent resumes on every message, so it remembers the ones before until New session starts another, and its history button lists the sessions to carry one on. The agent runs in the session's own folder, .workspaces/<id> under the projects folder, with the project in front, if there is one, as a folder it may write: boards go in that project's canvases, not in the working directory.

State and persistence

The document lives in the browser's IndexedDB under PERSISTENCE_KEY in the app's src/App.tsx. A board's identity is its path key, so renaming a folder or a file orphans that board's shapes; Force refresh rebuilds them. Ordinary layout drift is what refresh is for, not a persistence-key bump.

Working on the canvas app itself

Only when changing the app, not when using it, and in a checkout of the repo: sp root, run inside it, prints it.

bashCopy code
cd "$(sp root)/canvas"
bun run lint && bun run test && bun run build

Then, in a fresh browser session: each board page loads with its frames, headings and captions; the frames stay independently selectable; double-clicking a board zooms it to fill the window; Force refresh rebuilds a board cleanly.

Everything the canvas needs a server for lives in canvas/server/: sp.ts answers /__sp and /board for one project, projects.ts mounts one of it per project at /p/<name>/ and makes or opens projects, the Vite dev server mounts that for working on the app, and main.ts mounts it in front of dist as the server sp start and the app run. Board discovery is boards.ts, served as /__sp/index.json and fetched by the page before it loads, not an import.meta.glob. Bump PERSISTENCE_KEY only when a change would leave existing documents inconsistent with the code, such as a shape's props changing shape; a bump discards every persisted hand-drawn annotation.