Phantra AI: Model Prompting
Chooses Phantra video models from the catalogue, checks length, frame and credit fit, and guides model-aware scene wording and quoting.
Model prompting
One scene, in the studio's shape, renders on any model in the catalogue. Every model is handed the same scene text, with the studio's own continuity line in front of each shot; you never write a model's own syntax. What changes with the model is what it can take (length, frame) and what it does with the words (sound, one long take). Read those off the catalogue, never from memory. Reading the catalogue spends nothing; a render waits on a quote and a yes.
Read the catalogue
video_models: every model with the lengths in seconds it takes, the frames (resolutions), the credits a short clip costs at each frame, and the cheaper draft frame where it has one. These numbers are for choosing;quoteprices a scene exactly.recommend_modelwith the job in words ("a cheap test", "a final spot with sound", "one 20-second take"),seconds,budget_credits, anddrafttrue to prefer a draft tier. Each pick carries one reason and its credits at that length; the reasons also say what the person kept before. Repeat the reasons; add nothing about a model the catalogue did not say.- Check the fit: the frame must be one the model lists, and a whole-scene
durationone of its lengths; either off the list is refused at the quote, never rounded. A timed scene is priced shot by shot, each window fitted UP to the model's nearest legal length (a 3-second shot on a model whose shortest clip is 6 seconds is priced as 6), so keep shots at or above the model's shortest length to pay only for what is on screen.
What follows from what the catalogue says
- Sound. When a pick's reason says it generates sound with the picture, the sound line is heard. Write it concrete and sparse, still after "No background music.": "Only diegetic sound: the lid's scrape, one floorboard." Never ask for a voice saying words the person did not write. Keep the line for every model; what a model does with it is the model's.
- Length and cuts. A model whose lengths run to 20 or 30 seconds can take
a single continuous window: a scene written as one take renders whole at
its full length (
quotewithdurationprices that), so fewer cuts. A model whose lengths stop near 10 seconds renders a ten-shot scene one shot at a time, but cannot take a 20-second single take. Default to cuts of 2 to 10 seconds either way; one take only when the person asked for one. - Frame and draft. The frame multiplies the price. A draft frame (a
cheaper one, where
video_modelslists it) is for choosing a take; it is finished in the studio.set_preferenceswithvideo_modelanddraft_firsttrue makes aquotegiven no renderer choice price that model's draft frame; aquotepassedmodelorframeignores the preference, so for one scene pass the draft frame asframeyourself.
Choose, then approve
recommend_model; say the top pick, its reason and its credits, and the runner-up when the reasons differ.- The person picks.
set_preferenceswithvideo_modelfor the workspace's default, or passmodel(andframe) toquotefor this one scene. - Write the scene as write-scene says. The words do not change by model beyond the two points above.
quotewithmodelandframe. Say the price from the quote (each shot's seconds and credits, the keyframes, the balance) and WAIT for the person's yes in this chat. Only thenapprovewithwhat="clip", the quote'stokensand the sameprovider,modelandframe(anddurationfor a one-take scene);jobuntil it is done,scenefor the clips (price-and-render). A short balance answersinsufficient_credits: start nothing, and say credits are added in the studio.
A worked exchange, with the names and numbers taken from the tools' replies:
recommend_model(what="a final client spot with sound, two people in a
kitchen", seconds=10, budget_credits=300)
-> picks, each with a reason and credits for 10s at its default frame
Say: "For a spot with sound the catalogue ranks <first pick> first: <its
reason>, <its credits> credits for 10 seconds at <its frame>. <Runner-up>
is <its reason> at <its credits>. Which one? I can set it as your default."
quote(scene_id=411, model="ltx2.3", frame="1080p")
-> clip: one 6s shot with its credits and one token; keyframes: 1 still; the balance
Say: "On LTX 2.3 at 1080p that is one 6-second shot for <its credits> (its
still would be <each> more); you have <balance>. Render it?" On a yes:
approve(scene_id=411, what="clip", tokens=[that token], provider="fal",
model="ltx2.3", frame="1080p", duration=6), then job until done.
Do not
- Never state a rule the catalogue did not give you: a model's prompt syntax, a word limit, a camera tag, a negative-prompt field. Every model is handed the same scene text.
- Never quote credits per second from memory; read
video_modelsorquote. - Never pick for the person: rank, say why, and let them choose. A changed
model or frame re-prices the scene, so
quoteagain and approve only with the new quote's tokens and that same renderer choice. - Never call
approvebefore the person has said yes to the quoted price in this chat, never on their behalf, and never twice for one quote. - Never pass a length or frame a model does not list, and never promise a model will match a render made on another one.
- Downloads
- Security audit
- Pass
- Last updated
- 2w ago
- Current version
- v1.0.0
- License
- MIT-0
