Comic Creator

Kids 9-13 who can't draw but have stories make real comic books: they write the plot and dialogue, AI only draws what they direct. A printed book they made, Screen time that creates, not consumes.

Comment

Inspiration

My daughter was trying to sketch a cartoon like image and I generated one with her in a comic strip format, she got very excited and was puzzled on how AI was able to do that. That was the seed, then we sat down and discussed what she could do and decided to embark on the comic creation app idea. I hope to develop and refine it more but wanted to share the story.

Comic Creator is designed for children around ages 9–13. Its core promise is simple: screen time that produces a story the child made instead of content the child only consumed, while opening their eyes to what can be done with AI and their imagination.

What it does

Comic Creator guides a child through five steps:

  1. Create and approve an original hero.
  2. Choose and refine a visual style.
  3. Write the Setup, Problem, Big Moment, and Ending.
  4. Direct each panel, including the scene, action, dialogue, and captions.
  5. Present the finished comic and export it as a printable PDF.

The child writes every plot event, character fact, caption, and line of dialogue. AI can illustrate those directions, but it never writes or rewrites the story.

Artwork is never silently replaced. Every generated image becomes a candidate that the child can approve, dismiss, or redirect. Dialogue and captions remain exact, editable word boxes by default.

Children can add panels to any story beat, create comics of up to 16 panels, pass the pen between two writers on the same device, reopen a saved project after restarting the app, and present the finished comic four panels per page.

A bundled Sample mode demonstrates the complete editing, presentation, and PDF workflow without requiring an API key or making a paid request.

How I built it

The browser interface uses React, TypeScript, and Vite. An Express server manages OpenAI requests, validation, image handling, PDF generation, and local filesystem persistence. Shared Zod schemas protect the boundary between the browser, server, stored projects, and model responses.

The OpenAI models have deliberately separated roles:

  • GPT-5.6 Luna converts child-authored visual facts into constrained rendering choices such as framing, lighting, palette, and focus. It also powers an optional Story Coach that returns only a validated missing-element signal. The interface converts that signal into one fixed, neutral question.
  • GPT Image 2 creates the hero reference and reference-guided panel illustrations.
  • OpenAI Moderation checks visual-generation requests before they reach the image model.

API credentials remain on the local server. Projects and generated images remain on the local filesystem. The prototype has no accounts, analytics, cloud database, public sharing, photo uploads, or remote collaboration.

How I used Codex

Codex acted as an engineering and review partner throughout Build Week. It helped turn the product direction into checkpointed plans, write tests before behavioral changes, build strict schemas and provider boundaries, inspect browser and PDF output, diagnose integration failures, and coordinate independent code-review passes.

Codex accelerated implementation, but the consequential product decisions remained human decisions: the child-as-author rule, target age group, four-beat structure, local-only architecture, explicit artwork approval, editable lettering, fixed-question coaching boundary, Education category, and privacy constraints.

Challenges

The hardest challenge was preserving authorship while still making AI genuinely useful. A general storytelling assistant could easily take over the most meaningful creative decisions. Constraining GPT-5.6 to rendering choices and one classified coaching signal kept the child in control.

Character continuity, generation latency, moderation failures, exact lettering, and safe recovery also required careful handling. Generated artwork is treated as a candidate rather than immediately replacing approved work. Failed requests preserve the existing artwork and authored words.

Another challenge was maintaining one consistent composition across the editor, presentation view, and PDF. The final implementation preserves the complete square artwork and uses the same normalized coordinates for editable word boxes everywhere.

Accomplishments

I am proud that the project supports the complete journey from an idea to a tangible printable comic while retaining every authored word.

The final build includes local persistence, project recovery, reference-guided artwork, non-destructive revisions, multi-page comics, presentation mode, same-device coauthoring, Sample mode, and PDF export.

The deterministic verification suite contains 291 passing tests across the domain, server, client, storage, model contracts, failure recovery, responsive behavior, and PDF output, together with strict typechecking and a production build.

What's next

The next step would be supervised usability testing with children, parents, and educators. A production version would also require dedicated legal, privacy, parental-consent, abuse-prevention, and data-retention work.

Refining the style prompts and also having a way to nudge the kids to prompt more properly.

Potential product improvements include stronger character consistency, a recent-project gallery, my summer vacation-like thematic templates, additional print layouts, and carefully designed family or classroom workflows. Educational outcomes remain hypotheses to test rather than claims made by this prototype.

Built With

Share this project:

Updates