OpenAI Build Week 2026 · Education

Your words are the magic pen.

Maliang is a six-panel comic studio that turns revision into visible consequence—while the child stays the sole author of every story word.

Watch the under-3-minute demo Explore the source

Ages 8–10
70 tests
Zero-cost judge path

THE 3-MINUTE ROUTE

0:00The promise
0:35The revision loop
1:30The OpenAI build
2:20Proof & tradeoffs

THE AI DRAWS.
THE CHILD WRITES.

A finished Maliang comic panel showing a child with a lantern waving to a crescent moon outside a cave — REAL PRODUCT RENDER · THE MOON DOOR · PANEL 6

00:30 / THE IDEA

Most AI makes the picture the reward. Maliang makes revision the reward.

A child writes. Maliang turns supported story facts into an evidence-bound render contract. New words change what the picture is allowed to support—and a resolved writing challenge can earn a craft card.

WRITE

The child owns every word.

The editor stays exactly where they left it. Maliang never completes the sentence.

SEE

Words become evidence.

Each semantic fact passed to the renderer points back to an exact span in the story.

REVISE

The picture answers back.

The next render reflects a new contract, turning revision from homework into the mechanic.

Maliang app showing the Moon Door editor, six comic panels, and an optional writing-helper question — THE REAL APP · OPTIONAL WRITING HELPER

Child-facing questions are fixed, reviewed application copy—not model prose.

00:55 / THE LOOP

One question.
Their ideas.

Maliang never inserts text, suggests replacement wording, scores the page, or blocks finishing. Help starts with consent and returns control after one question.

01
Picture check
Is this how you pictured it?
02
Child chooses
Action, appearance, setting—or skip.
03
One question
Fixed, reviewed copy. Never generated prose.
04
Child edits
The story remains untouched until they type.
05
Compare
The child decides whether the new picture is closer.

01:25 / THE PAYOFF

The Moon Door grew one sentence at a time.

Selected real renders from the submitted app show the move from provisional sketch to a finished story.

A provisional pencil render of a child waiting in a cave beside a lantern — PROVISIONALMara waits.
A sparse sentence yields a contract with very little supported detail.

A rendered comic panel of Mara, a lantern, a cave, and a moon door — MORE EVIDENCEMara waits in the dark cave, holding her little lantern.
The cave and lantern are now backed by the child’s words.

The finished Moon Door story panel under a starry sky — STORY PAYOFFA finished six-panel comic
The visual reward belongs to the writer—not the coach.

01:45 / BUILT WITH OPENAI

OpenAI models inside a guarded pipeline.

Generative power sits behind deterministic contracts, schema validation, and revision-aware controls.

INPUT

Child’s words

The current panel text is the source of truth.

CODEX + GPT-5.6-TERRA

Structured reasoning

Safety, scene extraction, and diagnostic codes with schema-constrained output.

TRUSTED APP CODE

Evidence contract

Exact source spans, validation, stale-job rejection, and local dialogue composition.

GPT-IMAGE-2

Default illustration

The validated render contract becomes a 960×720 comic panel.

WHY CODEX MATTERS

Runtime engine and build partner.

Sandboxed Codex subprocesses run structured GPT-5.6 work. The submission also retains a GPT-5.6 Sol/Codex raster path for benchmark comparison; direct GPT-Image-2 is the demo default.

LOCAL-FIRST, HONESTLY

Encrypted at rest. Cloud-aware in use.

Project data and illustration artifacts are encrypted locally. Live generation still sends de-identified current-panel inputs and contracts to cloud models.

02:20 / PROOF, NOT PITCH

Tested, benchmarked, and candid about the gaps.

70/70tests passing

15test files

3renderers compared

0model calls in judge mode

Same-prompt comparison of raster, vector, and direct OpenAI API renderer paths with timing results — ONE SYNTHETIC FIXTURE · SAME PROMPT · THREE RENDER PATHS

ZERO-COST JUDGE PATH

Run the real UI without credentials or model spend.

npm install → npm test → npm run devOpen setup guide

MALIANG