Ad variations

A variant is a new scene copied from one that exists, with one thing changed, filed in the same project. The source is never touched, its stills and clips are never copied, and nothing is spent until the person has seen a variant's price from quote and said yes to it in the chat. Change ONE thing per variant, so what gets kept says something.

The order

1. Find the source: recent, search or project gives its scene_id; scene reads it in full (prompt, seconds, title, the timed shots) and gives only the COUNT of its photos; the refs themselves are what you passed, or project's references list, each naming the scenes it is on. Read it first.
2. Agree the variable with the person: hook, length, first shot, product photo or model. Draft the change, show it, get a yes.
3. duplicate_scene with the scene_id and ONLY the fields that change; the rest is copied. The reply gives the new scene_id and copied_from.
4. Note which is which (below), then quote each variant and say its price: keyframes (stills, each credits a still) and clip (its credits in all), beside the balance. Then WAIT. One yes covers exactly the quotes the person was shown and named; anything else needs its own yes.
5. On the yes, approve with that scene_id, what "clip", the tokens from its quote and the same provider, model and frame the quote was priced with (and its duration when the scene is one shot). Tokens are good for one hour; a changed choice is refused, never re-priced. what "keyframes" draws the stills instead, on a yes of its own. Then job with the returned id (wait up to 15) until it is done; scene shows the clips.
6. Later, what_works reads what was kept.

The five changes

• Hook (prompt): rewrite the first window's line only -- the first sentence said, or the first image -- in the person's own words, and keep every other window, the style block, the sound and the avoid list as is.
• Length (prompt + seconds): re-time the windows so they still run contiguously from 0 to the new total, each 2-10 s; drop or shorten a shot rather than squeeze them all; seconds is the last window's end.
• First shot (prompt): swap the opening window for a different image (the product before the face, or the reverse); the rest stays timed.
• Product photo (refs + roles): a different pack shot from elements as the first ref with role product (or a different face as subject). The first ref anchors the render, so this changes the picture most. A ref elements did not list is refused.
• Model: a scene carries no model; it is picked at the quote, so no copy is needed. recommend_model ranks models for the job in words; quote with model (and frame) prices the same scene on another one, and approve with those tokens and that same choice renders it there; set_preferences with video_model makes it the default wherever a call passes nothing (the quote then says so in preference; name it again at approve), and draft_first true prices the cheaper draft frame where a model has one. video_models lists them with credits per frame.

Telling variants apart

A copy is titled " (copy)" and cannot be renamed from here (a write_scene rewrite replaces a scene's text, never its title), so keep a short table in the chat as you go and say which scene_id is which:

Worked example

Source 412 is a 15 s to-camera ad with three windows and the person wants a shorter one. Re-time it: (0-4s) the open unchanged, (4-10s) the product shot two seconds shorter with the closing line moved to its end, the third window dropped. Then duplicate_scene with scene_id 412, the re-timed prompt and seconds 10: it gives scene 419. quote 419 with model ltx2.3 reads two shots, each with its credits, and their total (each window is fitted up to the model's shortest length), beside the balance. Say that and wait. On "yes", approve 419, what "clip", the two tokens, provider fal, model ltx2.3, frame 1080p; job until it is done; scene 419 shows the clip.

Reading the result

what_works says, per model and per preset, how many outputs were made and kept (downloaded, cut, starred, used again), how many tries a keeper takes, and what was posted with a scene behind it. When enough_data is false, say so rather than read a winner into two renders. scene on each variant shows which clips landed, with its studio link.

Do not

• Never change two variables in one copy and call it a test.
• Never call approve without a yes in this chat to that variant's own price, never approve on the person's behalf, and never carry one yes over to the next variant or a second render of the same one.
• When quote or approve answers insufficient_credits (with needs and available), stop and say so: credits are added in the studio.
• Never put a URL in refs (exactly as elements lists them), and never rewrite the source with write_scene to make a variant; copy it.