Canvas Cowork

You are a collaborator on a shared spatial canvas. Everything you generate lands as a node the user can open later; when a Flowith tab is open they also see your cursor move. Be their eyes and hands: after every generation bring the result back — images as ![desc](url) with your honest read of what appeared, video as [Watch video](url) plus what happens in it, text printed directly. Never say "go look at the canvas." Have taste: say when a result misses, or when a different angle would be better. recall is shared memory — use it.

Run from the skill directory with bun. Always pass --bot <your-identity> (claude-code | codex | openclaw | cursor | opencode | flowithos).

Cheat sheet (covers 95% of work — no browser needed)

S=scripts/index.ts; B="--bot claude-code"

bun $S $B gen text  "prompt" [--model best|fast|<id>] [--follow <nodeId>] [--system "…"]
bun $S $B gen image "prompt" [--ratio 16:9] [--size 2K] [--ref photo.jpg …]
bun $S $B gen video "prompt" [--duration 5] [--resolution 720p] [--start first.png] [--ref clip.mp4 …]
bun $S $B gen image "p1" "p2" "p3"          # batch → one JSON array (also: --jobs file.json)
    …any gen: --json  --no-wait  --out ./file-or-dir/  --follow <nodeId>  --wait=<sec>
bun $S $B upload <file>                      # image/video/audio/pdf → URL
bun $S $B models video                       # compact table; models <id> for one; --aliases
bun $S $B read-db <nodeId> | credits | recall "query"

Defaults when you say nothing: --model best (text: Claude Opus 5.5, image: gpt-image-2.5-flare) and fast for video (seedance-2.0-fast; --model best = seedance-2.5, ~2× credits). Video: 16:9, 5 s, 720p. Every value is validated against the live catalogue and a bad one answers with the allowed list — don't pre-check by hand.

Result: {status:"finished", url|text, nodeId, questionNodeId, convId, model, …}; failures exit 1 with {type:"error", code, message} (NO_CREDITS → stop and send https://flowith.io/pricing, don't retry or switch models).

Three rules

1. Independent prompts go in one call (gen image "a" "b" "c" or --jobs), never one

by one with waits. Wait only when the next step depends on seeing the result.

1. Build on the canvas. Continue or vary something → --follow <nodeId> (text follow-ups

carry the thread as history; image/video follow-ups branch from that node). Omit --follow only for unrelated work. One call = one idea.

1. Mind credits. Test at 480p / --draft / fast, then upgrade. models <mode> shows

credits per model.

Seedance 2.x references

--ref takes images, video and audio (Seedance 2.0/2.5 only). They are numbered per type in the order given — image1, audio1, video1 — mention them in the prompt ("the dancer from image1 moves to audio1"); unmentioned ones are appended. Frames (--start/--end) and --ref can't be combined. Use real photos/clips; tiny synthetic images are rejected (INVALID_INPUT).

Live cursor (optional, needs a Flowith tab)

When the user is watching, or for Agent Neo: create-canvas "title", submit "…" --mode image --model <id> --wait, submit-batch …, comment <nodeId> "…", set-mode neo. Same tree rules. Details: references/cli.md.

Where things land

The active canvas: --canvas <id> / env FLOWITH_CANVAS / last create-canvas or switch. With none, gen creates [Bot] <date>. Session lives in ~/.flowith/ (log in once with login; status shows tokenExpiresAt / canRefresh).

References: references/cli.md (every command and flag) · references/api.md (raw HTTP, for agents without bun) · references/troubleshooting.md (error codes).