How to write a comic script: panel-by-panel format, with a drawn example
Synopsis, page breakdown, PAGE and PANEL format, captions and balloons: how to write a comic script, with a full page written for this article and drawn with AI.

Frank HoubreFounder of Imaginode
8 min read
How to write a comic script
A comic script is broken into pages, then panels: for each panel you write what the reader sees, then the text on it (captions, balloons, sound effects).
The artist, or the AI drawing for you, can't read your mind. It reads panels. A good script tells it, panel after panel, where the camera is, who does what, and what goes in the balloons. Everything else is story, and you sort that out before you start.
Further down you'll find a complete five-panel page, written for this article and then drawn by Nano Banana Pro. The result is good, with one failure that taught us a writing rule.
From synopsis to script in three steps
Synopsis first, page breakdown second, panel-by-panel script last: each step answers a different question.
**The synopsis** fits in a few lines: who wants what, what stands in the way, how it ends. If you can't write it, go back to the story before you worry about format.
**The page breakdown** spreads the story across pages. One line per page is enough: "page 1, Ines delivers a parcel at night and finds out it's dated 1986." This step sets the pacing, and above all decides what lands at the bottom of the page, right before the reader turns it.
**The script** splits each page into panels. That's the document the artist works from, and it's what the rest of this article is about. Some writers call it a full script, as opposed to the looser "Marvel method" where the artist gets a plot summary and fills in the panels.
Comic script format, panel by panel
A PAGE heading with the panel count, then for each PANEL a description in the present tense and the lettering sorted by type.
There's no official standard, but most scripts look alike. At the top of each page: **PAGE ONE (5 panels)**. Giving the panel count up front lets the artist plan the grid before reading the details.
For each panel: **PANEL 1.** followed by the description. Shot first (wide, medium, close-up, high angle), then the setting if it changes, then the action. Present tense, plain sentences. "Ines reads the label on the parcel," not "Ines, worn out after a long day, wonders whether…" You draw the exhaustion: a soaked raincoat, slumped shoulders.
Lettering goes under the description, one line each, with its type in capitals: **CAPTION** for narration in a box, **CHARACTER NAME** for a balloon, **(OFF)** when the speaker isn't in the panel, **SFX** for a sound effect. The letterer, and the AI, then know what goes where.
Full example: "The Fourth-Floor Parcel"
Here's the script exactly as it was written and handed to the model: one page, five panels, one twist.
**PAGE ONE (5 panels)**
**PANEL 1.** Wide shot, night, rain. A narrow street in Paris, an old apartment building. A bike courier in a yellow raincoat stops at the door. Only one window is lit, on the fourth floor. CAPTION: Tuesday, 11:40 p.m.
**PANEL 2.** Medium shot. INES (mid twenties, short curly hair, soaked yellow raincoat) reads the label on a small parcel wrapped in brown paper and tied with string. INES: Fourth floor. No elevator, of course.
**PANEL 3.** High angle looking down a spiral staircase. Ines climbs, tiny in the frame, the parcel held tight against her. SFX (small, next to the parcel): tick… tick…
**PANEL 4.** Close-up. A door open a crack, chain on. In the shadow, an old man's eye. MR AUBRAC (OFF): You're late.
**PANEL 5.** Close shot. Ines, puzzled, looks at the postmark on the parcel: "MARCH 12, 1986". MR AUBRAC (OFF): By forty years.
Three things worth noticing. The caption sets the time in four words, no lead-in. Panel 3 has no dialogue: it gives the climb some time and slips in the small, worrying sound. The twist lands in the last panel, where the eye finishes the page.
How many panels per page, how many words per balloon
Somewhere between four and nine panels a page depending on the style, and rarely more than two sentences a balloon: past that, the page gets heavy.
These are habits of the trade, to adapt to your story. American comics often sit around five or six panels a page, European albums tend to pack more, webtoons use far fewer per screen, and manga swings a lot with the action. If you're starting out, four to six panels leave the art room to breathe.
On text, a balloon that runs past two sentences eats the panel. If a character has a lot to say, split it into two balloons, or two panels. Read your dialogue out loud: what sounds written sounds wrong, just like in an audio drama.
One last benchmark: a panel shows one moment. "Ines climbs the stairs, knocks and hands over the parcel" is three panels, not one. It's the most common mistake in first scripts, and AI pays dearly for it: the model tries to show everything in one image and gets none of it right.
Drawing the script with AI: what worked
With a character sheet and a prompt that follows the script panel by panel, Nano Banana Pro produced a readable page on the first try.
The method has two steps. First, Ines's character sheet: three views, front, three-quarter and back, on a plain background. It locks the raincoat, the hair, the delivery bag. More on that in our guide to keeping a character consistent.
Then the page: the prompt describes the grid (one wide panel on top, two rows of two), then each panel with its shot, its action and its exact text in quotes. The PAGE and PANEL format of the script copies over almost word for word. That's the point of a clean script: you use it twice.
Result: all five panels, in order, with the right shots and clean balloons. The French page, made from the same prompt with translated dialogue, came out just as clean, apart from a page number nobody asked for. Each image cost 19 credits, the price listed on the model's page.
The miss that makes a rule: an important detail gets its own panel
The only thing that failed is the postmark in panel 5, the key to the story: a tiny detail lost in a busy panel.
Look at the parcel in the last panel. The script asked for "MARCH 12, 1986". The model drew a string of unreadable marks in English, and "1606" in French. The whole twist rests on that date: without it, "By forty years" means nothing.
A human artist would have hit the same problem. A one-inch stamp in a panel that also has to fit a face and a balloon is something nobody reads. Comics solved this long ago with the insert panel: a close-up on the object alone.
So we rewrote the ending with one more panel, just the parcel and its postmark. First attempt, wrong in a new way: the model added balloons we never asked for ("THIS CAME FOR YOU."), because our style line mentioned hand-lettered balloons. Give the model two instructions that clash and it picks one. Second attempt, without that line and with "no text anywhere except the stamp": correct date, in both languages.
The rule for your script: **if the reader has to read a piece of information to understand the page, give it its own panel, or put it in the dialogue.** Image models write a title or a balloon well, and small text on an object much less well, as we saw when comparing which AI writes text in images best.
The same workflow in Imaginode
Imaginode's comic strip page builds this workflow on the canvas: a Reference node for the character, a Style node for the inking, one Image node per panel.
On the canvas, the character sheet goes into a Reference node. You then call it in every prompt with `@` and its name, so the face stays the same from panel to panel. A Style node holds the art direction, and each panel gets its own Image node, with whichever model you pick. The cost shows on the button before you generate.
You can open the comic strip workflow without an account to see how it's wired. If your story comes from a novel or a short story, character sheets from a book handles the first step from the PDF, and keeping the same character covers the reference part in detail.
What a page costs
At 19 credits an image with Nano Banana Pro, the character sheet and the page come to 38 credits, and every redone panel adds 19.
For the English version: Ines's sheet, the page, then two insert attempts, so four images and 76 credits. A credit is worth about one euro cent. Other models in the catalogue cost less per image, but Nano Banana Pro is a good pick when the page has lettering and a precise layout.
Generating panel by panel instead of a full page costs more credits but gives you more control: you redo the bad panel without touching the others. For a first comic, starting with full pages shows you quickly whether the script holds up.
If the comic needs to become a film or an anime
A script written panel by panel turns easily into a storyboard, then into animated shots.
The script you just wrote already looks a lot like a shot list: one framing, one action, one line per panel. If you ever want to animate your comic, none of that work is wasted. Our article on storyboarding before you shoot shows how a panel becomes a shot.
Your turn
Write one five-panel page before you think about a whole book: it's the best way to learn the format.
Take an idea that fits in one sentence, split it into five moments, keep the surprise for the last panel. Reread each panel and ask: can this be drawn? If the answer takes two images, cut.
Only then draw it, by hand or with AI. The script will show you straight away what's missing.
How we analysed this
The script for the page "The Fourth-Floor Parcel" was written for this article on November 1, 2026. It was drawn the same day with Nano Banana Pro, called through the Vercel AI Gateway, the provider Imaginode uses for this model, with a key from Frank's account. We did not go through the Imaginode interface: same model, same provider, different screen.
One generation per image, no reruns and no cherry-picking: one page in English, one in French, a character sheet, then two attempts at an insert panel per language. Every result is shown, failures included. The price of 19 credits per image comes from Nano Banana Pro's public page, checked the same day.
Imaginode is made by Frank Houbre, who writes this blog. The writing benchmarks (panels per page, words per balloon) are working habits of the trade, presented as such, not figures from a study.