| name | Grok Imagine Agent |
| description | Use when driving Grok Imagine Agent Mode on grok.com/imagine (project + infinite canvas) for a multi-shot film. Log every still/video ID from the first generate. Last frame of N is first of N+1. Ask Agent to trim and stitch joins. Spoken video goes through Agent chat with the still attached. Not Studio. Not x.com SuperGrok chat. |
Grok Imagine Agent
Drive Imagine Agent Mode at https://grok.com/imagine. URLs look like /imagine/agent/<id> — do not publish private IDs. Three panes: sidebar (projects / history), agent chat (Type to imagine), infinite canvas (dot grid + tools). Different from Studio (/imagine/studio) and SuperGrok chat on x.com/i/grok.
This is the web app in Chrome on this computer. Not a desktop app. Not the user's laptop.
Official Odyssey craft: Odyssey Challenge Cinematic Guide (2026-08-26). Contest deadline August 31. Do not contest-submit without a human yes.
ID ledger (from the first generate)
Every still and every video gets a canvas id the moment it exists. If you skip this, later edits guess thumbnails and joins break.
- The instant a still or video node appears, write a row: shot number, still UUID, video UUID, take letter, KEEP/CHANGE/DELETE.
- The ledger lives with the film (
CANVAS-MAP.md or equivalent). Update it before the next generate.
- Click by id, never by a guessed thumbnail.
- A new take gets a new video id. Do not overwrite the old row — add
take-b and keep take-a.
- Do not invent ids. Truncated ids stay marked INCOMPLETE. Missing ids stay NEED ID.
- Never publish
/imagine/agent/<id> or conversation query strings. Node UUIDs for shots are the working ids.
No generate without a ledger row.
Map the canvas, then edit cut by cut
- Zoom out until every shot node is visible (25–50%).
- Map node id → boarded shot. Write the map before you edit.
- Edit only CHANGE shots. Leave KEEP nodes alone.
- One cut at a time. Click that still id. Then left Agent chat with that still selected/attached.
- Still first (Speed → Quality upscale, pores/lines). Spoken video only after the still PASSes, via Agent chat + still. Node one-click Make Video skips the line.
- Download the new node. Write its new video id. Never overwrite
*_LOCK or keeper clips.
Last frame is first frame (do not skip)
Same camera / same location: the last frame of shot N is the first frame of shot N+1.
- Extract last frame (or use Agent "Extract Frame" at the end).
- That image is the start of the next still/video. Prompt: start on this exact frame, then play the new beat.
- New lock still only when camera angle or location actually changes.
- If start/end do not match, the join will slam. That is a skill miss, not a stitch trick.
Ask Agent to trim and stitch
Imagine Agent can trim and stitch. Ask it. Do not hard-concat locally as the first answer.
Name the join, the two video ids, and the transition:
- Same location: last-frame match, or a short dissolve. "Start shot N+1 on the last frame of video
<id>."
- Location change: a real cut or a named fade (e.g. 0.6s cave-mouth → ritual).
- Line ends → shades: trim the tail of the spoken take, then emerge from shadow into the next still — not a slam to OTS.
- Never close→wide on the same face. Hold size or cut away.
- Composer submits on every newline: stitch asks are one line.
Example one-liner: Stitch video <id-A> then video <id-B> start B on the last frame of A trim the tail of A so the join is continuous same cave no slam
ffmpeg pairwise xfade is a fallback after Agent stitch is refused or Picture FAILs the Agent join. It is not the default.
Official still → video
Do not write “cinematic” and stop. Name shot size, angle, move, lens.
- Aim.
- Speed Mode still.
- Quality Mode upscale (pores, fine lines).
- Animate with omni-reference.
- Prompt like a director: shot / doing / speech / remember. No music.
Two video paths (do not mix)
Spoken — Agent chat
Attach the selected canvas still. 1080p, 6/10/15s, Audio ON. Exact line. No extra words. Never into lens. NO MUSIC.
Silent — canvas tap on that node is fine
Audio OFF. One job.
While it runs
- Stay on the live agent tab. One Chrome. One job.
- Download when the node is done. Size >1MB. Log the new id.
Do not
- Use Studio if Agent + canvas is live.
- Hard-concat or local xfade before asking Agent to trim/stitch.
- Skip last-frame handoff on a same-location join.
- Generate without writing the still/video id.
- Fire a second generate while one is Generating.
- Publish private
/imagine/agent/<id> or conversation IDs.
- Invent scenes. Human pass locks keepers.
- Contest-submit without a human yes.