Token economy
The plugin cuts tokens in three places:
- what is loaded at all times: as little as possible;
- what Claude reads: JSON summaries, a contact sheet and targeted frames rather than raw output and dozens of images;
- what gets redone: only what changed is re-checked, and the project is not re-explored every session.
What the plugin does on its own
| Mechanism | Effect |
|---|---|
| Almost nothing loaded at all times | About 270 tokens loaded at all times: the description of the motion skill and those of the two agents. The eleven /umd:* commands cost nothing until you run them. In a umd project, the startup hook adds about 150 tokens of context. |
| References on demand | 56 guides of 800 to 2,300 tokens each, read one at a time, only when needed: for example the project's route, then its engine's guide. |
| Checks done in code | 29 of the 64 checks are decided by the CLI from FFmpeg measures, lint, licences and browser tests. The auditors only look at what needs eyes. |
| Contact sheet | Each render produces a 16-frame sheet: a single image to read. Claude then extracts only 2 to 6 targeted frames if a doubt remains. |
| Compact output | Renders, encodes and analyses write their full output to disk and return only a summary. In our test, 3 commands went from about 3,700 tokens of output to about 700. |
| Incremental audit by default | Only the checks that a changed file or a new render can affect are re-checked. If nothing changed, no auditor is started. |
Quick mode (/umd:audit quick) | Automatic measures only, no agent. |
| Local searches | umd styles and umd moves search indexes shipped with the plugin: zero tokens, no network. Only the chosen style is downloaded. |
Drafts (/umd:render draft) | Half resolution at 15 fps while iterating. |
| Cheaper model for the auditors | "audit_model": "sonnet" in .motion/project.json; your session stays on your usual model. |
Handoff (/umd:handoff) | A short note lets the next session resume without re-exploring. The startup hook points to it for 14 days. |
| A style card instead of the prompt | The chosen style prompt (1,200 to 7,700 tokens) is read once, summed up in 200 words in BRIEF.md, and only that summary is read again, by the critic and by the build. |
| Grouped auditors | A group with 3 checks or fewer is given to another group's agent: an agent for two criteria would cost as much as one for fifteen. |
Transparency
- Warned before spending. Before each step,
/umd:gostates its token cost: measured on this project's earlier runs when there are some, estimated otherwise (agents launched, images read). An expensive step waits for your go-ahead, with the cheaper options: in practice the build (about 120 to 290 k tokens for 30 s, critique included) and the first full audit (about 110 to 250 k). In.motion/project.json,"cost_confirm": "medium"asks more often,"never"never; with atoken_budget, any step that would exceed it waits for your go-ahead. /umd:tokensreads Claude Code's local transcripts for this folder, subagents included. For each/umd:*command it shows input, output, cache-write and cache-read tokens. It adds the one change that would save the most on your project.- With a
token_budgetin.motion/project.json, it shows the share of the budget used. - The audit report also shows the tokens saved by compact output.
By kind of piece
| Piece | Typical cost | To spend less |
|---|---|---|
| Short (logo, micro-interaction, 2 to 10 s) | Low: few beats, one sheet is enough | /umd:audit quick while iterating, a full audit before delivery |
| Medium (promo, presentation, short, 15 to 90 s) | Medium: the storyboard and drafts dominate | Validate the look on stills before animating, 2 draft rounds, auditors on Sonnet |
Long (narrative film, explainer of several minutes, director route) | High: many scenes, assets and reviews | Work sequence by sequence (/umd:build 1-6), audits by group, /umd:handoff at the end of each session |