Intuit — 2025

Intuit For Education

Scholarship explainer timed by GPT-4 and assembled in After Effects by script

A machine-timed explainer for Intuit's scholarship-types RFP — GPT-4 acting as animation director, ExtendScript building the After Effects timeline from its JSON, boards trained on Intuit's own illustration style.

Timed Scenes

120.82s

End Timecode

Image Reference Set

The hard part

GPT-4 acted as animation director — gapless scene timings and camera moves as JSON, and ExtendScript built the After Effects timeline from it

Client

Intuit

Agency

MALKA Media

Role

Creative Technologist AI Pipeline Architecture / Animation Direction

Timeline

2025

Automated

The timeline built itself

Intuit for Education · Financial Aid Navigation Pathway — Scholarship Types · RFP — AI-assisted pipeline

The delivered film

After Effects 2023 with Final_JSON_Sequence open: twenty-eight layers — every artboard placed and timed by the script in a staircase across two minutes, the voiceover underneath, no keyframes — **Final_JSON_Sequence, as the script left it**1024×1820 · 2:00:20 at 24fps · every artboard placed and timed by the script — the staircase is the edit, and no keyframe in it was set by hand. The delivered film beside it is this comp, rendered.

v6 · 2:06 · 9:16 · fourteen boards, timed by a model, assembled by a script

runtime: 2:06
boards: 14
scenes: 14, gapless
hand-keyframed: 0 frames

The tool came firstHundreds of scripts were possible, so the first thing built was a script analyzer — and if it was already reading the script, it could draw it too.

script in: 360 words
scenes out: 14
cut to: 2:00.8
holds: 7.1–16.4s
camera moves: punch in ×4 · out ×4 · pans ×2 · held ×4

STEP 1／9 One .txt, pasted. No shot list, no prompts, no scene numbers — 360 words of prose is the entire brief.

One beat, all the way through — The same fifteen seconds at four stages — nothing here is paraphrased.

01the script section

Jake — Athletic & Community Scholarship (0:15–0:45). Jake, a stylized baseball player, catches a ball. The scene transitions to show a split screen: one side shows a baseball field, the other a community center icon. A stylized check with “Tuition” appears, partially covered by two smaller checks labeled “Athletic Scholarship” and “Community Scholarship.”

What a human wrote.
02what spaCy pulled

JakeAthleticCommunityScholarshipVisualbaseballplayercatchesballscenetransitionssplitscreenbaseballfieldcommunitycentericonstylizedcheckTuitionappears JakeAthletic & Community ScholarshipVisualJakestylized baseball playerballscenesplit screenbaseball fieldcommunity center iconstylized checkTuition

45 keywords by part of speech — nouns, verbs, adjectives — and 25 noun chunks kept as phrases. A sample of each is shown.
03the prompt it wrote

Create a 2D illustration in the style of Intuit brand videos (simple shapes, clean lines, bright colors, hand-drawn style character). Scene description: Jake, a stylized baseball player, is catching a baseball. Visual elements based on keywords and phrases: Jake, stylized baseball player, baseball, Jake, athletic, baseball, player, catches. Character style should be similar to the example video provided (simple, illustrative, 2D). Color palette should adhere to Intuit visual guidelines.

Composed from the extraction above. Nobody typed this.
04what came back

Artboard 03, in the brand line, from a prompt the pipeline wrote itself.

Open the build record Eleven days, from an empty file to a comp that builds itself — including the three app screens the drawing above stands in for.

the analyzer

**script_analyzer.py**spaCy for entities, a Hugging Face sentiment pipeline, and scene segmentation that splits on Scene: markers and falls back to double newlines.

**gradio_ui.py**analyze_and_generate: the analyzer's scenes, a swappable fal model list read from JSON, and a cache so a re-run costs nothing.

the app

**the first version**Gemini, one button, no style control. The model picker and the split Analyze / Generate came after.

**the tool, running**The real script pasted in, fal-ai + stable-diffusion-v3.5 selected, scenes generating underneath.

**scenes as a table**Every scene as a row — description and visual elements — and a button that turns the table back into a script.

the style search

**one style prompt, every frame**A style prompt applied across all images, so fourteen boards come back looking like one hand drew them.

**style, tested**The same five scenes under different style prompts — the search for the Intuit line before the LoRA existed.

the trained line

**the trigger word**Intuitive-Hand-Drawn-Line-Art — the token that summons the brand's own line.

**training runs**Two LoRA trainings, minutes apart, both completed.

the boards

**fourteen boards**What came back, in the brand line, each with the description that produced it.

the assembly

**the comp, assembled**Final_JSON_Sequence: 1024x1820, 2:02, every board placed by the ExtendScript from the model's JSON.

**the staircase**The timeline a script builds — layers stepping in sequence, not one keyframe set by hand.

One beat, all the way throughThe same fifteen seconds at four stages — script, extraction, prompt, board.

sweep to scrub

The self-assembling editOne explainer, from script to cut, where the timeline is machine-built.

Three students, three scholarship mixes. A script, a voice read, and fourteen one-line board descriptions — that is everything human-authored that enters the pipeline.

the inputs, in full

First, meet Jake, a high-school baseball player who just earned an athletic scholarship to play at a Division II university… Artboard 03 — Jake looking straight to camera, in a baseball uniform, with a high school baseball game in the background.