token-saver: tokprep

The exact command for this machine is in your system prompt ("Local media/file prep on this machine: …"). On macOS that is the tokprep script, also linked at ~/.local/bin/tokprep; on Windows it is python "<plugin>\bin\tokprep".

Each command uses the Apple tool on macOS (Vision, Spotlight, sips, PDFKit). Elsewhere it uses the Windows alternative when the machine has one: Windows OCR and face detection (built in), Windows Search or Everything (es.exe), rembg for cutouts, pdftotext, and ffmpeg for sheets. If a command answers NOT AVAILABLE on this machine, do the task directly, as you would without tokprep: read the image or PDF yourself, use your own search. The hooks let a second Read of the same file through. Each command does the work on the Mac and prints a short result, or writes summary.md plus at most two labelled sheets. Read the summary first. Open a sheet only when the summary can't answer the question, and open a raw file only when the sheet can't either.

Making an asset: always inside a job

When you edit a video or make any asset for the person (a cutout, a key, a gif, an edit, a render), keep the intermediates inside a job. That way they disappear once the final is approved.

1. tokprep job start "<what you are making>" --dest "<folder the person wants the final in>"
2. Put every intermediate in $(tokprep job dir): tokprep's own outputs land there by themselves, and pass -o "$(tokprep job dir)" to your own ffmpeg/sips/python commands. Anything temporary you had to write elsewhere, register with tokprep job track <path>.
3. Show the person the candidate and ask whether it is the final version.
4. When they approve it as final: tokprep job deliver <file> [--to <folder>] [--name <final name>], then tokprep job done. Tell them where the final is and how much space was freed. The mod compacts the conversation after that turn, so the job's iterations stop taking up context.
5. If they want changes, keep iterating in the same job. If they drop the asset entirely, use tokprep job abandon.

tokprep job done refuses to run before a final is delivered, never deletes a delivered final, and only deletes things the job made: its work folder, plus tracked files newer than the job and outside top-level or destination folders. Use --dry-run to see the list first. tokprep clean --days 7 removes old prep summaries and sheets made outside jobs.

Rules:

• Outputs: final assets go wherever the person says (job deliver --to); never put intermediates in their folders. tokprep never overwrites a file; it adds -1, -2 to the name.
• Cutouts: check the coverage numbers. 0% means no subject was found; above 80% means the background was kept. Look at the before/after sheet only if those look wrong. --mode person is for people only; subject also catches objects and dark silhouettes.
• Chroma key: a green halo means raise --similarity (0.15–0.2); holes in the subject mean lower it (0.08).
• Formats: transparent video comes out as ProRes 4444 .mov. WebP is not available in this ffmpeg build: use AVIF or JPG.

Code: graphify first

For questions about a codebase, use graphify, not tokprep. When graphify-out/graph.json exists, run graphify query "<question>" before grepping or reading source, graphify path "<A>" "<B>" for relationships, graphify explain "<concept>" for one concept, and graphify update . after code edits. No graph yet: suggest /graphify . once. If graphify is not found, run tokprep graphify setup (the plugin does this once, after it is installed).