Working with the AI

You don't prompt Filmara. You ask it for work, the way you'd ask a collaborator — "write a scene for every beat", "get images for my characters" — and it answers with a plan you can read before anything runs. This page collects the asks that matter, in plain words.

Three things are true for every ask on this page:

• Your words are enough. Every ask here is written the way a person says it. If a natural phrasing doesn't do what it obviously should, that's a bug on our side — not a vocabulary lesson for you.
• Nothing runs until you press Start. The plan card shows every step, what the plan read of your project, and — when steps generate images — the credit ceiling it's asking to spend. Start authorizes exactly that, no more.
• Every documented ask is tested. Each one on this page runs against Filmara continuously, asserting the exact plan shape you see described. If the product changes, this page fails our own build until it's true again.

Create a story from scratch

What you say

Create this story end to end — the structure, the main characters, and a scene for every beat.

You don't need this exact sentence. "Build the whole story", "take this from idea to first draft" — anything that asks for the full arc works. What matters is that you've talked about the story first: give it a brief, even two sentences, so there's something to build from.

What happens

Filmara answers with a plan card, not with pages of content. Read it — it's the contract. For a full-arc ask you'll see the structure laid out first, then the characters, then one scene per beat, each as its own step. The card also shows where the project stands as the plan read it, and flags anything that looks off — a plan that ignores what already exists gets caught here, before anything runs.

Press Start. The run performs one step at a time and narrates itself in the task's progress log. The structure lands first — later steps bind to the real beats it creates, not to guesses.

When it's done

The chat tells you plainly: how many steps applied and what actually changed — "scenes 0→11" is measured, not assumed. Your beats each hold a scene, your cast exists with profiles, and everything the run wrote sits above a "Before Task" checkpoint in Versions, so the whole thing is one step to undo.

Worth knowing

Words-only work costs nothing. If the story has open questions the plan would quietly settle — genre, ending, whose story it is — the card lists them; starting the task authorizes those calls, or you can settle them in conversation first.

Write a scene for every beat

What you say

Write one scene for every beat.

Or "fill the structure", or "write the scenes for Act 2" — the ask names the coverage you want. This recipe assumes your structure already exists (beats on the board); if it doesn't, use Create a story from scratch instead and the outline comes first automatically.

What happens

The plan card shows one scene step per empty beat — a beat that already has a scene is left alone, never overwritten. Each step targets its beat by name and id, copied from your real structure.

On Start, scenes land one at a time, each into its own beat — you can watch the count climb in the progress log. If you type mid-run, the run finishes its current step and pauses to hear you; nothing is killed halfway.

When it's done

Every beat holds its scene, in story order, under the beat you'd expect — check the structure view. The done message reports the measured change ("11 applied · scenes 0→11"), and the "Before Task" checkpoint in Versions undoes the whole run if you hate it.

Worth knowing

The run honors what exists: established characters appear in the scenes, recorded story decisions are conformed to, and a beat that gained a scene since the plan was drafted is skipped honestly rather than duplicated.

Create — or deepen — your characters

What you say

Create the main characters for this story.

or, when they already exist:

Develop the existing characters — deepen all of them.

You don't have to know which word is which — Filmara checks your cast. Ask it to "create" someone who already exists and it deepens them instead; you never get a duplicate CHLOE.

What happens

One step per character on the plan card. A create introduces someone new, grounded in your brief and story so far; a develop enriches an existing profile — fills the gaps, keeps what's already true. Each step runs as its own turn, so each character gets full attention rather than a batch paragraph.

When it's done

The cast page holds them all — profile, role, appearance, backstory. The done message counts what changed, and every profile is a record you can edit by hand afterward; nothing is locked.

Worth knowing

Profiles are the foundation the visual work builds on: avatars and shot renders read them for faces, wardrobe, and presence. Deepen a character before generating their images and the images get better for free.

Get images for your cast

Spends credits: ~5 credits per character.

What you say

Generate avatars for every character.

"Generate images for the characters", "give my cast profile pictures" — same recipe. Filmara resolves people by name, so this works for characters created seconds ago in the same task or ones you made weeks back.

What happens

One image step per character on the plan card — and because these steps spend credits, the card shows the ceiling before you start. Start authorizes up to that amount, never more. Each image is composed from the character's own profile: appearance, age, presence — which is why deepening profiles first pays off.

When it's done

Each character has their image generated and assigned — it's their avatar on the cast page and everywhere their face shows up, not a loose file in a folder. The done message accounts for every one.

Worth knowing

Images render one at a time; if your credit balance runs out mid-run, the run pauses honestly at that step and tells you — everything already generated stays assigned.

Plan the shots for a scene — and render them

Spends credits: ~15 credits per scene render.

What you say

Propose the shots for the wedding scene and render them.

Name the scene the way you think of it — "the wedding scene", "the scene where they're discovered" — Filmara matches it against your script. Ask for just the proposal ("propose shots for…") and the render step is left out; "and render them" adds it.

What happens

Two steps on the card: a shot proposal — the scene broken into shots with framing, movement, and a composed prompt each — then a render step that generates the images for every proposed shot. The render spends credits, so the card shows the ceiling; Start authorizes it.

The proposal step runs first and queues the shots; the render step then works through the queue one image at a time.

When it's done

The scene's shots sit on the Storyboard with their rendered stills — framing notes intact, ready to reorder, regenerate, or hang on the timeline. The done message counts shots proposed and images rendered.

Worth knowing

One proposal covers the whole scene — camera angles, framing, movement, lighting, and composition arrive together in that single step. Asking for them one by one ("figure out the camera angle, then the lighting…") still lands as one proposal; they're facets of the same shot plan, not separate jobs.

Filmara knows which scenes already have shots. A broad ask — "propose the shots the script still needs" — plans proposals for the scenes that have none and leaves shotted scenes alone. Wanting more coverage of a scene that's already shotted is fine too; just say so — "more coverage of the wedding scene" — and the new shots stack alongside the existing ones.

Renders inherit your project's Look — set it first (era, palette, tone) and every shot lands in the same world. Rendering without a Look works, but consistency is what the Look is for. Animating a shot into video is its own ask, once a render exists.

Direct your scenes

What you say

Direct my scenes.

Or name one — "direct the wedding scene". Direction is its own ask: it stages the scene, it doesn't shoot it. Say "direct the wedding scene, then propose its shots" and both land as one task, direction first.

What happens

One direction step per scene on the card. Each step reads the scene the way it lives on the Storyboard — the ordered action and dialogue cards, with whatever direction and shots already cover them — and proposes staging beats: how the characters and the environment behave, how what happens happens, written the way direction is given on set. A beat directs one of the scene's moments and its card lands at that moment's column, in the director's own lane — riding the script, never pushing its cards apart. Verbose is welcome — a full sentence or two on how the moment behaves is good direction.

In chat, @direct proposes one scene's beats as a card: each beat shows where it lands, you tick the ones you want, and Accept places them. A beat whose spot no longer exists in the scene grays out with the reason instead of landing somewhere wrong.

When it's done

The accepted beats sit on the Storyboard as yellow direction cards in the director's lane, each stacked on the column of the moment it directs — the director's own track riding the script. Each card remembers its moment and when it was directed: rewrite the scene and a card whose moment survives stays with it, while one that loses its moment goes visibly unplaced — never deleted, never guessed somewhere wrong. The done message counts the beats placed.

Worth knowing

Direction comes before shots — it's the first step of the directing chain (direction → shots → renders). When a task both directs and shoots a scene, the direction step runs first, and the shot proposal reads the placed direction: a standalone direction beat is a natural moment for coverage.

Direction is staging, never camera work. Shot sizes, lenses, movement and transitions belong to the shot proposal; a direction beat that says "close-up on her hands" is a shot note in the wrong register — say "her hands stop moving" and let the shots decide the framing.

Filmara knows which scenes already have direction. A broad ask — "direct the scenes that still need direction" — stages the scenes that have none and leaves directed scenes alone. Re-directing a scene is fine too; just say so, and the new beats stack alongside the existing ones (pruning is yours, on the Storyboard).

Get the camera out of your script

What you say

Clean the camera direction out of my script.

Or point at one scene — "doctor the wedding scene". Any phrasing that asks to remove camera language from the script works: "get 'the camera pans' out of my action lines", "de-lens act one".

What happens

One doctor step on the card, sweeping the whole script in one pass (point at a scene instead and the step targets just that scene). The step doesn't rewrite anything — a detector finds the lens language in the action lines ("the camera pans over…", "we see that…", "close on…") and removes exactly those words, keeping the world they described. "The camera pans up to the sky, showing the sun shining down" becomes "The sun shines down." Blocking and staging are never touched — "he walks slowly, scanning the ground" is story, and story stays.

The same detector marks lens language in the editor as you write: a dotted underline under the phrase, nothing more. It never blocks you — you may want "we see" in a draft on purpose. The agent is held to the harder rule: it doesn't write camera direction at all.

When it's done

The done message counts the action lines cleaned and the scenes they came from. A scene with no camera language skips itself silently — an already clean script reports exactly that. Your script reads as story — what happens in the world, with no one behind a camera.

Worth knowing

The camera isn't lost — it was never supposed to live in the script. Framing and movement are authored on the Storyboard, where they belong: "direct my scenes" stages the action, "propose shots" frames it. Doctor first, then direct — the chain rebuilds the lens deliberately instead of leaving it scattered through the prose.

Heal a broken structure board

What you say

Fix my structure — the beats lost their scenes.

Or any phrasing that says the structure board itself is hurt: "my structure looks broken", "an act won't render", "heal my structure", "scenes fell out of their beats". The agent also knows on its own — when the structure is wounded, its assessment of the project says so.

What happens

One doctor step on the card — never a step per beat or act. The step doesn't redesign anything: the app diagnoses the structure document, and any scene that fell out of its beats is adopted back into the beat it sits under in the script's running order. Placement is by position, never by story intent — if a rescued scene belongs in a different beat, drag it there on the now-working board. Scenes are never touched, beats are never deleted; only the membership records heal.

The same heal is one click away without the agent: when the Structure lens detects the wound it names it in a banner — "Act 4 couldn't be placed — 1 scene fell out of its beats" — with the heal action inline. Both doors run the identical step, so either way the work leaves a task record and a "Before Task" checkpoint in Versions.

When it's done

The done message reports the findings — which scenes were re-adopted and into which beats, and anything else the diagnosis found (dead references dropped, duplicate claims released). A healthy structure reports exactly that ("structure already sound") and changes nothing — asking twice is safe.

Worth knowing

The doctor restores validity, not story. Redesigning the structure — new acts, different beats, a fresh outline — is "lay out my structure" territory (a beat_outline step), and moving scenes between beats is a drag on the Structure board. Wounds mostly enter through old-format version restores; after any restore the app checks and points you here if something came back hurt.