moda-video-footage
<!-- moda:banner -->
PREREQUISITE — load once per session (step-0, write contract,
free/metered map). Already loaded? Skip ahead. If you cannot load it, the
non-negotiables: run
before anything;
before creating; writes that pin a revision use your last read's — on
, re-read and retry once (it heals); send the canvas link the
moment it exists; stuck or failed?
— free and fast,
never guess.
<!-- /moda:banner -->
What this lane is
The user has REAL clips on disk — a shoot, a trip, screen recordings — and
wants a cut: a film, a reel, a highlight. Analysis stays local (Rushes, the
optional toolkit riding
, indexes and searches the footage on
this machine and uploads nothing); only the selects that make the cut are
uploaded, and the deliverable is always BOTH the editable Moda timeline and
the rendered mp4. Generating new footage is moda-video-clip's; motion over
designs and type is moda-video-motion's; this lane cuts what already exists.
Read references/footage.md before the first command — it is the whole
workflow with its invariants; references/otio.md carries the timeline
interchange contract the endgame rides. What follows is the shape, not a
substitute for either.
Recipe — folder to film
- Toolkit check:
moda rushes doctor --json
. Typed
teaches the fix — (it states
what it downloads first; sizes are multi-GiB). Can't install? Degrade
honestly: pick selects by eye with your harness's own tools and skip to
step 5 — say the selection was manual, never pretend an index ran.
- Index:
moda rushes ingest FOOTAGE --project INDEX
→
moda rushes analyze INDEX
→ moda rushes status INDEX --json
. The
project directory is persistent memory — keep it, reuse it across
revisions, never wipe it to re-ask a question.
- Brief + probes: one short exchange for target duration, tone,
priorities. Then probe search quality with a few
moda rushes search INDEX "…" --top-k 8 --json
queries and
moda rushes show INDEX mom_… --json
— look at contact sheets and
transcript evidence; captions can hallucinate and rank is not confidence.
- Choose:
moda rushes highlight INDEX --target-minutes N
for the
candidate pool (~2.5× the target, one per duplicate cluster — material,
not a film), then moda rushes select INDEX mom_… --name … --reason … --json
in play order, with exact rational trims via when a moment
needs shaving. Reason at duplicate-cluster level; consume returned ids.
- Extract + upload:
moda rushes extract INDEX sel_… --handles 2s --out SELECTS --json
(or upload whole originals when the folder is a few small
files), each select, rewrite the media map's
entries to with the REAL returned asset ids.
- Timeline: author and dry-run it first —
moda edit import CANVAS_REF cut-v001.otio --media-map media-map.json --dry-run
— against a canvas you created (moda canvas create --category animation
); then apply the same command with in place of
(replace mode clears the existing timeline, so it refuses
unconfirmed). Read the fidelity report; surface every dropped/approximated
item. New versions are new files — v002, v003 — never an overwrite of the
only cut.
- Draft, look, iterate:
moda export CANVAS_REF --format mp4 --scope main_edit -o draft.mp4
, sample frames, judge against the brief.
Small fixes are operations against a fresh read;
a recut is the next version.
- Deliver: the live canvas link FIRST — the human always gets an
editable timeline — then the mp4 and the version file when the
user wants an interchange copy for Resolve/Premiere.
Everything here is unmetered: indexing, uploads, import, export and
iteration are free, so iterate on the cut as much as the piece needs. A
music bed is an upload away (an audio track on the same timeline); a
generated score is moda-audio's lane.
Boundaries stated plainly
- Rushes never uploads, never calls cloud inference, and needs no Moda
account; Moda's cloud renders and exports and never reads local disk. You
are the bridge, and selected media is the only cargo.
- The user chose uploads of SELECTED media — never bulk-upload the folder.
- Supported transitions and visual tracks import and render. Unknown effects,
unpaired overlaps and resource limits decline by name in ;
relay them honestly. Never claim an unsupported edit exported without loss.
- An audio-only timeline does not export as mp4 — keep a visual clip on the
cut.
- Full inference is Linux x86_64 and Apple Silicon; elsewhere use the
media-only profile or the manual degraded path, and say which ran.
Errors
Any typed error → moda-core's recovery reference.
/
: run
, then re-run the SAME
command.
: the fidelity items name each offender —
fix them (upload + rewrite
entries; re-reference intermediates
with their own timings), never invent asset ids.
on an
edit: re-read and retry once — it heals.
Make it recurring
The weekly cut from a growing folder → moda-automate; the finished film's
social variants → moda-social; a designed cover still → moda-image.
See also: moda-video — the family fork · moda-video-motion — type and
motion over the cut · moda-core — contract, recovery, everything Moda can do.
References
| Doc | Load when |
|---|
| references/footage.md | ALWAYS before the first rushes command — the workflow, its caching/timing invariants, degraded paths |
| references/otio.md | before authoring or importing a timeline — media maps, fidelity report, exact-time discipline, OTIO vs edit ops |
| references/export.md | rendering the draft and final — scopes, ceilings, warnings |
| references/reading-and-verifying.md | the review loop; verifying without vision |