Skip to content
Log inSign up free

Prompting guide

How to prompt Plum

How to prompt Plum, Mage's full-feature video model: a quick start, a director's brief template, and expert tips for references, dialogue, sound, and edits.

Try Plum FastPlum model page

A man flexes in a gym mirror as his tank top melts into chrome knight armor
Made with PlumA man flexes in a gym mirror as his tank top melts into chrome knight armor

Quick start

Your first great Plum prompt

Everything a beginner needs in one minute: what to know first, the rules that matter most, a template to fill in, and prompts to copy.

What to know first

Plum reads a director's brief. Before it renders, it expands your words into a detailed script of its own, so your job is to say plainly what the camera sees and what we hear, not to invent a script format. A full brief has four parts: what each attached picture or sound is for, one sentence on the whole clip, the moments in order with their times, and the rules for the whole clip. Plum makes sound with every clip, so the brief always covers the soundtrack.

The rules that matter most

  1. Brief it like a director

    Describe what is on screen and on the soundtrack in plain, visible words: "rain beads on the window, a kettle starts to whistle". Plum follows direct description far better than metaphor or mood.

  2. Give every input a job

    Open with one line per attachment: "@jun is the character reference; keep his face and round glasses. @harbor is the environment reference." An input with no job is guesswork.

  3. Split the length into time ranges

    Write "0–4 s:", "4–8 s:" and so on, adding up exactly to the length you set. In each range: the shot size, what we see, the action, any words, the sounds.

  4. Settle the music

    Describe the score you want (the instruments, the tempo, where it builds), or end with "Do not add any additional background music." Asking for both confuses it.

  5. Draft on Plum Max, finish on Plum

    Both read the same brief. Plum Max is the faster model; switch to Plum when the keeper needs its higher resolution.

Fill in this template

Replace each part in square brackets with your own idea, and leave out a part you do not need.

Template
Inputs: [what each picture is for].
Concept: [one sentence on the whole clip].
0–4 s: [first shot].
4–8 s: [second shot].
Requirements: [look and light]. No [what must not appear]. [Music, or no added music].
The template, filled in
Inputs: @rosa is the character reference; keep her face, braided grey hair and green apron. @bakery is the environment reference.
Concept: before sunrise in a small bakery, Rosa (image 1) pulls the first loaves from the oven, warm and quiet, handheld, standard cuts.
0–4 s: Wide shot. Rosa opens the oven door and heat blurs the air; she slides out a tray of loaves on a wooden peel. The oven door creaks, the crusts crackle as they cool.
4–8 s: Cut to a close-up of Rosa. She breaks a loaf open, breathes in the steam, and says softly: "Perfect."
Requirements: warm light from the oven, flour dust in the air, shallow depth of field. No subtitles, captions or watermarks. Do not add any additional background music.
  • [what each picture is for] One job per attachment, and what must not change: @rosa is the character reference; keep her face and green apron. Leave the whole line out when nothing is attached.
  • [one sentence on the whole clip] Who, where, what happens, the style, the camera, and how it is cut: before sunrise in a small bakery, Rosa pulls the first loaves from the oven, warm and quiet, handheld, standard cuts.
  • [first shot] Shot size, what we see, the action, any line in quotes, the sounds.
  • [second shot] Start with "Cut to" and the new framing, and say who is on screen.
  • [look and light] Light the camera can see, the palette, the texture: warm light from the oven, flour dust in the air.
  • [what must not appear] Exclusions, as plain items: subtitles, captions or watermarks.
  • [Music, or no added music] A described score, or "Do not add any additional background music."

Prompts to start from

  • A vertical drama from text alone

    Concept: a vertical live-action standoff in a cramped noodle shop at night, natural short-form drama performances, shot mainly in medium close-ups with shot/reverse-shot cutting.
0–4 s: Medium close-up. A young cook in a stained white apron slams a ladle on the counter and glares across it. He says, voice tight: "You said you'd pay me Friday."
4–8 s: Cut to a medium close-up of the owner, a heavy-set man in a cardigan, slowly folding his newspaper. He answers flatly: "It is Friday. Tomorrow."
8–10 s: Cut back to the cook, jaw clenched, steam rising between them from a pot.
Requirements: warm tungsten light, red paper lanterns soft in the background, shallow depth of field. Pots bubble and a fan hums throughout. No subtitles, watermarks or stickers. Do not add any additional background music.

    Why it works: With nothing attached, the brief describes both men, the place, and the style itself, and the ranges add up to ten seconds. Set the length to 10 seconds and a 9:16 aspect ratio.

  • From an opening to an ending frame

    The opening frame shows a paper boat at the edge of a rain puddle; the ending frame shows it caught against a storm drain grate. Animate the boat from the opening frame to the ending frame in one natural, continuous shot with no cuts: a gust pushes it off, it spins once, rides the stream along the curb and lodges against the grate. Low angle at street level, rain pattering, water gurgling into the drain.

    Why it works: It says which picture opens and which ends, asks for one continuous shot, and spends the rest on the path between them and its sound.

  • A short brief for a brand film

    A premium outdoor-gear film. @image1: mood, setting and film grain. @image2: the climber. @image3: the backpack. @image4: the ending logo. Story: on a windy ridge at sunrise, the climber pulls the backpack tight, scrambles up the last rocks and stands at the summit as the sun breaks through. Tone cool and confident; edit fast, with beat-synced cuts.

    Why it works: A short ask can fit in a few sentences with the jobs inline. Upload the four pictures in that order, and set the length to 15 seconds and a 16:9 aspect ratio.

Advanced

Advanced Plum prompting

Techniques for more control once the basics work.

Plum or Plum Max

Plum Max is the faster model and the default; Plum renders the highest resolution. The same brief works on both, so try the timing on Plum Max and switch models for the keeper.

On Plum, a First Frame or a Last Frame can stand alone. On Plum Max, a Last Frame needs a First Frame beside it.

The director's brief, part by part

Inputs: number each attachment as Plum receives it, give it one job, and name the features that must not change. Concept: one sentence that sets up the whole clip, tying each element to its input in parentheses: "a courier (image 1) races across a rooftop market (image 2)".

The time ranges: split the length into ranges that add up to it exactly, and in each give the shot size, what is visible, the camera move, the action, any words, and the effects. Requirements: the rules that hold for the whole clip (the look, the palette, the light, how text and characters behave), then the exclusions as "No …" and "Do not …" lines.

  • Plum cuts between shots by default. For one unbroken shot, write "one continuous take" in the Concept and describe the action as one paragraph, with no time ranges.
  • With nothing attached, Plum invents everything, so a text-only brief needs at least the subject's appearance, the setting, the action, and the style.
  • Plan text-only coverage in layers: a wide shot for the space, a medium shot for the action, a close-up for the detail.
  • Keep the brief to a few short paragraphs. A brief that runs on as one block loses its parts.

Camera, coverage, and editing

Say how each shot is framed and how the clip is cut. Shot/reverse-shot suits a conversation, facial close-ups build tension, and a wide shot establishes each new place. Name the editing style in the Concept so Plum cuts the way you would.

At every cut, name the new framing and who is on screen: "Cut to a medium close-up of the courier (image 1)".

  • Other treatments that work: aerial shot, slow motion, handheld, subtle push-ins and pull-backs, dissolves and fades.
WriteWhat it does
shot/reverse-shotAlternating views of two people talking, each over or toward the other.
locked-offA camera that does not move.
fast lateral trackingThe camera races sideways alongside the action.
smash cutAn abrupt cut from one moment to a very different one.
jump cutA cut within the same shot that skips ahead in time.
beat-synced cutsCuts that land on the beat of the music.
white-flash cutA brief flash of white between two shots.
hidden cutA cut concealed behind something passing close to the lens.
J-cut / L-cutThe next shot's sound starts early, or this shot's sound runs on after the cut: smooth joins in dialogue.

Look, light, and texture

Name light the camera can actually see: "a cool, soft front-side key", "backlight keeps the foreground dark and the distance bright", "shallow depth of field with snow as soft bokeh". Then the palette ("low-saturation cold blue, ink green and charcoal") and the texture ("heavy grain", "halftone dots").

Write what the camera sees, not what the scene feels like. "Her hands shake as she folds the letter" works; "a moment heavy with regret" leaves Plum guessing.

References, Characters, and frames

Mention saved Characters and References by @handle and images you upload as @image1, @image2. Mage gives each a number before Plum reads it: your uploads first, then Characters in the order you first mention them, then References. @jun becomes "character from image 2", or "character from image 2 with voice from [Audio 1]" when he has a saved voice. Saved audio clips are numbered [Audio 1] and up, ahead of the Characters' voices.

Plum wants every input named and given a job, so after the Inputs line, call each one by name with its number in parentheses where it helps: "Jun (image 2) pours tea". Jobs an input can have: character, object, or environment; opening or ending frame; voice; storyboard (one picture per shot); style or composition; motion or camera motion; audio to reuse, whole or a stated stretch.

  • For a style or mood picture, add "Reference only the stated qualities; do not reproduce the image itself."
  • A described face changes from shot to shot. To keep a person, attach them, as a Character or an upload, and name that input as the character reference.
  • Frames and references are separate modes: one clip uses frames or references, never both. A saved audio clip needs a picture beside it.
  • An @ followed by a word is always read as a handle, and one that matches nothing stops the generation. In dialogue, spell an email address out in words: "jun at harbor dot com".

A Character, a location, and a product

Inputs: @jun is the character reference; keep his face, round glasses and orange windbreaker. @harbor is the environment reference. @image1 is the object reference: the matte black thermos; keep its shape and white logo exactly.
Concept: at dawn on the harbor, Jun (image 2) pours tea from the thermos (image 1) and toasts the fishing boats heading out, cinematic and calm, a slow push-in, standard cuts.
0–4 s: Wide shot. Jun sits on a bollard as boats slide past in the mist. He unscrews the thermos and steam curls up. Gulls cry, engines thrum in the distance.
4–8 s: Cut to a medium close-up of Jun. He raises the cup toward the boats and says quietly: "Bring them all home." The camera pushes in slowly.
Requirements: cold blue dawn light with a warm glow on his face from the east. No on-screen text. Do not add any additional background music.

Why it works: The upload comes first, so the thermos is image 1 and Jun image 2. Each input has one job and its fixed features, and his saved voice speaks the line. Set the length to 8 seconds.

Dialogue, voices, and sound

Write every spoken line word for word, in quotes, after its speaker: The courier (image 1) says, breathless: "Last drop of the night." Fit the words to the time range they sit in: crowding a short range with dialogue is the main cause of broken lip sync. When a line runs across a cut, say so: "she continues speaking from the previous shot".

Mark a speaker the camera cannot see as off-screen or voiceover, and when a cut reveals them, say it: "Cut to a close-up of the old man, revealing him as the voice."

  • A Character with a saved voice brings it along. For any other voice clip, write "Her voice timbre follows [Audio 1]", and do not repeat the clip's words when you want only its sound.
  • Dialogue holds best in Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Write each line in the language it should be spoken in.
  • Put each sound effect in the time range where it happens, and steady ambience (rain, traffic, a fan) in the Requirements.
  • Keep voice clips short, and say which part of a clip matters.

Text on screen

Quote every word that should appear: The following English text appears on screen: "NIGHT MARKET". Or place it in the scene: The phone screen displays the title "Track Order" and a button labeled "Open Map".

  • Show one main piece of text per shot, placed clear of faces and eyes.
  • When the words matter, add "All text is spelled and rendered correctly, with no garbled characters."
  • Attach a brand logo as an image with the job "ending logo". To keep brands out, add "Do not reproduce brands, logos or text from the inputs."
  • When nothing should be written, add "No captions, stickers, watermarks or on-screen text."

Editing a video you add

Add a clip in the Video setting and call it video 1. A plain instruction per change works best, saying what changes and what stays: "In video 1, replace the cat with a dog." "Replace the view outside the window in video 1 with @image1." "Change the final line from '…' to '…'." "Relight video 1 from midday to blue dusk."

An edit takes the clip and a reference picture or two, without frames or audio. With a voiced Character in the prompt, turn off Use Character Voices.

  • A video can lend only its motion: "She performs the dance from video 1; her appearance follows @image1."

Swap a jacket, keep everything else

In video 1, replace the grey suit jacket on the man with the red leather jacket from @image1, keeping its zips and collar. Keep his face, his movements, the camera path, the street and the original audio unchanged.

Why it works: One change, the details of the new piece to keep, and everything else, sound included, named as unchanged.

Common mistakes

Avoid this, do this instead

  • Avoid: The whole brief as one unbroken paragraph.

    A courier on a rooftop market at night jumping between stalls and the camera follows and there is music and she says something about the last delivery and it should look cool.

    Instead: Inputs, concept, timed moments, requirements.

    Concept: a courier races across a rooftop market at night, live-action, fast lateral tracking, fast cuts.
0–3 s: Wide shot. She vaults a fruit stall as lanterns swing; crates clatter.
3–6 s: Cut to a medium close-up. She lands, grins, and says, breathless: "Last drop of the night."
Requirements: neon and lantern light, light rain. No captions or watermarks. A driving electronic beat that cuts out on her line.

  • Avoid: An input with no stated job.

    Instead: "@image2 is the environment reference."

  • Avoid: Music asked for and ruled out in the same brief.

    Instead: One decision about the score, or a rule for each part.

  • Avoid: "One continuous take" alongside several time ranges.

    Instead: One continuous paragraph of action, with no ranges.

  • Avoid: A face expected to stay the same with nothing attached.

    Instead: The person attached and named as the character reference.

  • Avoid: A text-only brief of one short line.

    Instead: At least the appearance, the setting, the action, and the style.

  • Avoid: Time ranges that run past the length you set.

    Instead: Ranges that add up to the length exactly.

  • Avoid: Metaphor and mood words.

    Instead: What the camera sees and the microphone hears.

FAQ

Plum prompting questions

How long should a Plum prompt be?
A short ask fits in two or three sentences with the jobs inline. A full brief runs to a few short paragraphs: inputs, a one-sentence concept, the timed moments, and the requirements. Plum expands it into its own script, so clarity matters more than length.
How do I keep a character consistent in Plum?
Attach the person, as a saved Character or an upload, and name it as the character reference in the Inputs line with the features to keep. At every cut, say which character is on screen. A description alone will not hold a face.
Can Plum do text?
Yes, short text: quote every word, say where it appears, keep one main piece per shot, and add "All text is spelled and rendered correctly, with no garbled characters." For a logo, attach it as an image with the job "ending logo".
Does Plum support negative prompts?
There is no negative prompt setting. Put exclusions in the Requirements as "No …" and "Do not …" lines.
How do I make a Plum clip without music?
Plum makes sound with every clip. End the brief with "Do not add any additional background music." and describe the ambience and effects you do want.
Which languages can Plum speak?
Dialogue holds best in Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Write each line in its own language.
Should I use Plum or Plum Max?
Plum Max is faster and the default; Plum reaches the highest resolution. They read the same brief, so draft on Plum Max and switch for the final when you need the extra resolution.
Can Plum edit a video?
Yes. Add the clip in the Video setting, call it video 1, and write one plain instruction per change, saying what changes and what stays the same.

Make your first video with Plum Fast

New accounts get 300 free Gems, enough to try Plum Fast. No install, no credit card.

Try Plum Fast