An AI rehearsal partner for actors. Feed it a script, tell it who you're playing, and it runs the scene with you: it performs the other roles, waits for your lines, and keeps track of where you are — so you can rehearse anywhere, any time, without needing another person.
This is the public case-study repo for Scene Partner (working name ScenePal). The production code is private; this repo documents the product, the design, and the AI-assisted development workflow behind it, with mock data and public-safe examples.
Left: script setup — paste or upload a scene, pick your role. Right: rehearsal mode — the app performs the other characters and cues you when it's your line.
Actors rehearse lines constantly, and the standard options are bad: begging a friend to read flat lines off a page, covering the page with your hand, or expensive coaching time spent on rote repetition. A rehearsal partner should be available at 11pm before an audition. This is a real workflow problem from working theater/film actors in my circle, not a demo in search of a use case.
flowchart LR
A["Script in<br/>(paste / upload)"] --> B["Role parsing<br/>character + line map"]
B --> C["Pick your role"]
C --> D["Rehearsal loop"]
D --> E["App performs<br/>other roles"]
E --> F["Waits and listens<br/>for YOUR line"]
F --> G["Tracks position,<br/>cues on stumble"]
G --> D
- Script parsing turns raw script text into a character/line map (mock example in
mock/). - The rehearsal loop alternates between performing other characters' lines and waiting for the user's line.
- Position tracking means you can stumble, repeat, or jump back without the app losing the thread.
- Mobile-first: rehearsal happens in hallways, cars, and green rooms, not at desks.
- Latency beats polish: a rehearsal partner that lags kills the scene's rhythm. Fast local state, minimal round trips in the loop.
- The actor owns the pace: the app follows the human, never the reverse. Same human-in-control principle as the rest of my automation work.
Scene Partner was built end to end with an AI-assisted workflow I use across all my products:
- Problem framing: real user complaint -> scoped product definition with explicit non-goals.
- Agent-assisted build: coding agents (Claude Code / Codex) work from context packs and scoped tasks; I review every change. The human owns taste and product judgment.
- Tight verify loops: each feature lands with a manual rehearsal test against a real scene before moving on.
- Crystallized lessons: recurring fixes get written into the workspace's operating notes so the next feature doesn't repeat them.
What I'd improve next: richer stumble detection, per-character voice options, and a shared rehearsal mode for scene partners in different locations.
mock/sample-scene.txt— a public-domain scene formatted the way the parser expects.mock/parsed-scene.json— what the role/line map looks like after parsing.- Product screenshots in
assets/.
I'm Thaddeus Arndt — AI workflow builder and music/audio technologist. More: github.com/thaddeusarndt · preballin.com/portfolio