Appointment television for your homework. ReRun turns notes into an interactive retro-TV episode: it teaches before it asks, pauses for retrieval, and adapts a retry after a miss.
The no-key demo is fully playable offline: five original, voiced pilots, bundled MP3 narration, original scene/host art, captions, deterministic option shuffle, and local saved-episode storage. The media bundle is intentionally checked in so a judge can experience the full loop without an API key or a network call.
ReRun was pair-programmed with Codex end to end during OpenAI Build Week, and live episodes are written by GPT-5.6 as validated structured output. See Built with Codex and GPT-5.6 and the human-reviewed audit trail in CODEX_LOG.md.
Power on, select any show from the lineup, complete the recap, watch the teaching steps, intentionally miss a question, then select Rewind & retry. The corrective branch returns to the authored simpler question after one cue; a second miss reveals the answer. The remote’s pause/play, rewind, fast-forward, AUTO, captions, mute, and ratings controls all operate real player state. Use the TV guide to change shows.
Each pilot occupies its own channel (CH 04–08). The Act 2 retrieval check and finale are both part of the aired path.
Paste notes or add text, PDF, DOCX, PPTX, CSV, photo, audio, or video files. Extracted text is placed back into the editable notes box before generation. Audio/video must be under 25 MB; .mov should be exported as mp4 or webm. On Vercel, large uploads can exceed platform request limits, so the full multi-file demo is best shown locally.
Configured live generation uses GPT-5.6 and an original-theme guardrail. GPT-5.6 access can be restricted by project; if your key cannot use it, set OPENAI_MODEL to a model available to your project. The production Please Stand By screen remains visible while the episode is written; after validation, an illustration screen streams original per-scene art previews and final images. A failed art request falls back to the built-in scene treatment instead of blocking playback.
Saved generated episodes appear in Your recent episodes for seven days. After their artwork is ready, Download video captures the show in the browser and sends the WebM recording to the local server for MP4 conversion. Video export needs a supported browser and a live key for its narrated scenes. It uses ffmpeg when available (set FFMPEG_PATH when it is not on the server PATH), but safely downloads the valid WebM recording when no encoder is installed. The bundled no-key pilots remain the reliable judge path.
Prerequisite: Node.js 18.18 or later (Node.js 20 LTS recommended).
npm install
npm run devOpen http://localhost:3000. No .env file is needed for the bundled demo. Add OPENAI_API_KEY only for live generation, extraction, narration, or scene art.
Create a local .env file only if you want the live AI features. The bundled
demo needs none of these variables.
# Required for every live OpenAI feature. Keep this server-only; never use a
# NEXT_PUBLIC_ prefix for it.
OPENAI_API_KEY=sk-...
# Optional model overrides
OPENAI_MODEL=gpt-5.6
OPENAI_INGEST_MODEL=gpt-4.1-mini
OPENAI_IMAGE_MODEL=gpt-image-2
OPENAI_TTS_MODEL=gpt-4o-mini-tts
OPENAI_TRANSCRIBE_MODEL=gpt-4o-mini-transcribe
# Optional: path to ffmpeg for MP4 conversion during browser video export.
# Omit it when ffmpeg is already on PATH.
FFMPEG_PATH=/usr/local/bin/ffmpegOPENAI_API_KEY enables live episode generation, document/image/audio
extraction, narration, and scene art. The other OpenAI variables are optional
overrides; leave them unset to use the defaults shown above. For Netlify,
add the same variables in Project configuration → Environment variables
and redeploy. Do not commit .env or expose the API key in a browser-facing
NEXT_PUBLIC_* variable.
npm run typecheck
npm run test
npm run buildReRun was built in one week for OpenAI Build Week, pair-programmed with Codex throughout. The full, human-reviewed audit trail — every milestone, what Codex contributed, and how each change was verified — is in CODEX_LOG.md.
How we collaborated with Codex. We worked contract-first: architecture, the EpisodeSpec schema, scene-authoring rules, and the test plan were drafted on day one as a build contract, and Codex implemented against them. Every milestone was human-reviewed before it was logged, with the verification performed (typecheck, tests, production build, browser checks) recorded per entry.
Where Codex accelerated the workflow. Codex turned the day-one build contract into an integrated vertical slice: the CRT player (a ~1,200-line React state machine covering shots, question takeovers, branch reactions, and reduced-motion support), the ~600-line Zod EpisodeSpec validator, the branching beat engine, all five API routes (episode, ingest, scene-image, TTS, video export), the demo-audio prerender pipeline, and the 25-test suite. That meant the complete P0 loop — boot flow, recap, branching, commercial review, remote, and responsive UI — was runnable on the first build day. It also shortened iteration: Codex carried out three focused CRT framing and viewport rework passes over the next two days while we repeatedly typechecked, built, and browser-tested the product. Once we chose a no-key judge path, Codex implemented the offline catalog, deterministic option shuffle, audio manifest, and media fallbacks that made the decision real.
Key product, engineering, and design decisions were ours. A judge-first MVP scope; bundling all 46 MB of original media (268 narration MP3s, 32 art plates) so the full demo needs no key and no network; the original-art-only guardrail in every image prompt; teach-before-ask enforced as a schema requirement rather than a prompt suggestion; and the positioning against passive audio study formats. Codex executed these decisions; it did not make them.
How GPT-5.6 contributes to the final result. GPT-5.6 (Responses API, structured outputs) is the runtime author: it writes each live episode against the EpisodeSpec schema, so every generated show is guaranteed to carry typed teach steps before its first question and a corrective retry branch per concept — pedagogy enforced at validation time. Around it, gpt-image-2 streams per-scene art, gpt-4o-mini-tts voices the host, and omni-moderation-latest screens generation inputs.
ReRun is study support, not grading or a substitute for instruction. Generated content is constrained to supplied material and uses original, classroom-safe themes; the app does not imitate named characters, artists, or studios.