|Docs

Cut a trailer or teaser from your film

You get a trailer and a teaser from your film, also called a promo, a sizzle or a cutdown. Your agent finds the strongest moments and lines, builds the cut to the music so the big moment lands on the drop, adds title cards, freeze frames and hits, and keeps each length on its own branch.

Use invideo for this when you want the agent to do the first pass from the whole film: it reads the dialogue, the sound and the music for you, and every cut stays on a timeline you can change.

Quickest path

Set up your agent first with invideo for AI agents. Then give it this prompt.

Cut a trailer and a teaser from my film ./final/the-mill.mov with themusic ./trailer-track.wav. Use the best lines of dialogue. Land thetitle on the drop. Put the trailer and the teaser on their own branches.Price the transcript first.

In a chat app, you drag the film and the music into the project. In a terminal, you name the files and the agent uploads them.

What you need

A finished film or a locked cut works best, because the trailer uses final sound.

  • The film. The final export or the locked cut.
  • The music. A track with a build and a drop, or a mood for stock or generated music.
  • The title and the cards. The film's title, the tagline, the release date or festival laurels.
  • The lengths. For example a full trailer and a much shorter teaser.
  • What not to show. The ending, a twist, a cameo.

Steps

The agent works on branches and tells you the timecodes of every moment it picks.

  1. Bring in the film and the music

    The agent makes a timeline with new_timeline. For a large film file, it calls start_upload, PUTs the bytes, then calls add_media with the upload_id. Expected result: a sequence URL and a media bin id for the film and the music.

  2. Find the moments

    The agent reads the film three ways.

    1. For the lines, the agent prices audio.transcript with dry_run on the film's sequence URL from add_media and tells you the price. For a price of 1 credit or more it waits for your yes. Then it starts the transcript with max_credits set to that price, and calls again after retry_after_s for the words.
    2. It runs audio.onsets and audio.content for the loud and quiet moments.
    3. It splits the film into its scenes with video.sceneSplit and looks at each with video.frames.

    Expected result: a list of lines and moments with timecodes. You strike anything that gives away the ending.

  3. Map the music

    The agent runs audio.sections and audio.markers on the music to find the build and the drop, and audio.onsets for the hits. Expected result: the times of the drop and each hit, turned into frames.

  4. Build the cut to the music

    The agent creates a branch called trailer with timeline_history. It places the music on A2, then each moment on V1 with place_media, so the cuts land on the hits and the biggest moment lands on the drop. It places each line with linked_audio true. Expected result: a new state_token and a cut that follows the music.

  5. Add cards, freezes and hits

    The agent adds title cards with add_text between shots, and holds a frame on a hit with retime kind freeze. For extra impact, it makes a sound with the generation pattern.

    1. It runs find_models with the sound in words, picks one, and reads its inputs with get_model. A good pick is elevenlabs_sfx. Your agent can search for others with find_models.
    2. It calls generate with a prompt for the hit or the riser and dry_run: true, and tells you the price.
    3. For a price of 1 credit or more it waits for your yes. Then it calls generate with max_credits set to that price and an idempotency_key, reads the sequence URL with read when the result lands, and places the sound on the effects track A3.

    Expected result: cards and hits on the beats you planned.

  6. Make the second length on its own branch

    From the finished trailer, the agent creates a teaser branch with timeline_history action create_branch. On it, it keeps the hook, one line and the title, and removes the rest with edit_time_range mode close_gap. Expected result: two branch URLs, one per length.

  7. Check, then export each

    On each branch, the agent runs capture at each card and the drop, and check_timeline. It checks that no card has a typo and the title lands on the drop. Then it runs export on each branch's URL with target_lufs. Expected result: two export cards with status completed.

The same job from a terminal:

invideo edit new <project-url> --name "The Mill trailer" --fps 24invideo asset upload <edit-url> ./final/the-mill.mov ./trailer-track.wavinvideo media audio.transcript <film-sequence-url> --dry-runinvideo media audio.transcript <film-sequence-url> --max-credits <price>invideo media audio.onsets <edit-url> --asset m1invideo media audio.sections <edit-url> --asset m2invideo media audio.markers <edit-url> --asset m2invideo edit show <edit-url>invideo branch new <edit-url> --name trailer --token <state_token>invideo gen models <project-url> --query "cinematic boom hit sound"invideo gen model <project-url> --model elevenlabs_sfxinvideo gen run <page-url> --model elevenlabs_sfx --dry-run \  -i prompt="deep cinematic boom hit" -i duration=3invideo gen run <page-url> --model elevenlabs_sfx \  --max-credits <price> --idempotency-key trailer-hit-1 \  -i prompt="deep cinematic boom hit" -i duration=3invideo status <thread-url> --waitinvideo notebook show <thread-url>invideo edit show <trailer-branch-url>invideo edit change <trailer-branch-url> --timeline T1 \  --token <state_token> --operations @trailer.jsoninvideo edit show <trailer-branch-url>invideo branch new <trailer-branch-url> --name teaser \  --token <state_token>invideo edit show <teaser-branch-url>invideo edit change <teaser-branch-url> --timeline T1 \  --token <state_token> --operations @teaser.jsoninvideo edit capture <trailer-branch-url> --at 10,45,88,112 --out ./framesinvideo edit capture <teaser-branch-url> --at 2,10,24 --out ./framesinvideo edit export <trailer-branch-url> --format video --target-lufs -14invideo edit export <teaser-branch-url> --format video --target-lufs -14

Where the values come from: <project-url> from invideo project new, and <edit-url> from invideo edit new. <film-sequence-url> is the url that invideo asset upload returns for the film; m1 and m2 are the bin ids, which the reads that return at once take with --asset. <price> is the price each dry run returns, its estimated_credits, and <thread-url> is the url of the paid gen run. <page-url> comes from invideo notebook show <project-url>. <trailer-branch-url> and <teaser-branch-url> are the branch_url each branch new returns. <state_token> comes from the latest invideo edit show on the URL the next command runs on: <edit-url> for the first branch new, then each branch, read fresh after each change. A token from another branch, or from before a change, is stale. The files in --operations @trailer.json and --operations @teaser.json are bare lists of ops, like the operations list below.

The key MCP call lands a line before the drop and puts the title on the drop. This fence is the full MCP edit_timeline call. Each frame is the time in seconds times the frame rate.

{  "url": "<trailer-branch-url>",  "timeline_id": "T1",  "expected_state_token": "<state_token>",  "operations": [    { "op": "place_media", "asset": "m1", "track": "V1", "at_f": 2016,      "source_start_seconds": 3121.0, "source_end_seconds": 3125.0,      "mode": "overwrite", "linked_audio": true },    { "op": "add_text", "track": "V3", "start_f": 2112, "end_f": 2208,      "text": "THE MILL" }  ]}

The agent reads the new clip's id with get_timeline. In the next call, it sends retime kind freeze on it with at_f 2100 and duration_f 12 to hold the hit.

Cost and time

Edits and exports are free. Generation and transcripts cost credits. Paid stock costs a small credit charge, or one of your workspace's iStock licenses. Your agent runs a dry run first and tells you the price before it spends anything.

  • Costs credits: the film's transcript with audio.transcript, each sound from elevenlabs_sfx, and stock or generated music.
  • Free: the cut, cards, freezes, branches and both exports. audio.onsets, audio.content, audio.sections, audio.markers, video.sceneSplit and video.frames show their price in the dry run before anything starts.

The transcript of a feature takes a few minutes. The first pass of a trailer takes the agent under an hour. See Credits and what they cost for prices.

If something goes wrong

Most problems come from a drop that misses, or a cut that gives too much away.

SymptomFix
get_model or generate refuses the model as not foundModels retire. Run find_models with the job in words, use the best match, and tell the person which model you used.
The title misses the dropRead the drop time again from audio.markers, round it to the frame, and move the card with move_clip.
A line is cut mid-wordMove the edge a few frames out with trim_clip, then listen again.
The trailer gives away the endingTell the agent which scenes are off limits. It removes them. Ask it to add them to invideo_rules.md, so later cuts skip them too.
A silent scene has no moments pickedSilent scenes give no sound cues. The agent picks them by frames with video.frames.
The teaser changed the trailerIt was edited on the same branch. Restore the trailer from its checkpoint with timeline_history action restore.
audio.transcript is refused as not_foundIt got a bin id like m1. Pass the film's sequence URL instead.
edit_timeline refused with an old tokenSomeone edited first. Read with get_timeline, then rebuild the ops.
Avoid: exporting main for both lengthsInstead: export each branch's URL. The branch in the URL is what gets exported.
Avoid: calling export again to check on oneInstead: status or read the card's URL.

Prompts to try

Each prompt is a real variation of the job.

Make a short vertical teaser for Instagram from the trailer branch:the hook, the title on the drop, and the date. Captions on the line.
Give me a trailer built on dialogue and one built on music and imagesonly, each on its own branch. Tell me which lands harder.
Add a boom on every title card and a riser into the drop. Price thesounds before you make them.

When to use something else

Theatrical trailers often need a rating card and a mix to a cinema spec. Make the rating card from the rating body's own artwork, and mix to a cinema spec in a DAW like Pro Tools.

For a trailer music cue made to picture by a composer, cut the picture first, export it, and send it to the composer.