Repository documentation
Animated onboarding production brief
Story and visual direction
Audience: an engineer with an existing codebase who wants to try Literate AI without surrendering source ownership. Sam challenges a concrete consequence; Literate AI answers it, and the explanation changes what appears on screen. Their dialogue is collaborative, with dry skepticism rather than alternating sales narration. The useful outcome appears first: understand an onboarding plan before acknowledging a mutation.
Preserve the established project identities from ../assets/PROMPTS.md: Sam's
dark hair, beard, navy/amber engineering clothes, and LitAI's teal hair and jacket.
The new vector puppets have separate head, arm, eyes, eyebrows and mouth groups.
The first renderer used deterministic generic arm/nod cycles, fixed gaze and
procedural blinks/breathing. It did not realize purposeful acting; the user rejected
that animation. The replacement direction is specified in the
line-by-line performance score, including listener reactions,
props, eyelines and deliberate stillness. The directed renderer now implements that score with two-joint foreground arms,
phrase-cued poses and focus, gaze, expressions and explicit holds. Visual validation
belongs to the user; implementation and technical checks do not grant it. No portrait is edited, and no third-party
character artwork or voice likeness is used.
The browser companion and film share the same scene and recording clock. A separate animated clip for each spoken turn prevents clip-duration rounding from accumulating into lip/caption drift. The film's visible captions survive forwarding; the browser offers a static-character mode, normal controls and full transcript. No autoplay, unbounded loops or global keyboard interception are used.
Source ledger
| Claim or choice | Authority and limit |
|---|---|
| Exact release wheel; Python 3.11+, Git and an authenticated coding CLI | docs/user/getting-started.md, preparation section; this audition is not the complete installation walkthrough. |
| Planning is read-only; apply and acknowledgement are separate | src/literate_ai/cli/operator.py and litai onboard adopt --help; the retained source-run observation in evidence.json proves only its explicit scope. |
| Original source remains owned by the project | Existing onboarding/adoption authority and the unchanged two-file fixture. No build, baseline execution or acceptance is claimed. |
| Plan readiness does not establish release qualification | docs/user/framework-flow.md and the existing lifecycle/release gates. The video explicitly refuses that inference. |
| Hybrid Motion supports controlled object/SVG animations | Motion animate documentation; the optional package lock pins the actual runtime. |
| Magpie voice discovery and recorded synthesis | NVIDIA API example; exact voice IDs come from the observed catalog, not a guessed list. |
| Expressive stock voice styles | NVIDIA speech NIM documentation; voice/style availability must still be observed for the endpoint being used. |
codex-ios was inspected locally at revision
c3ad16c7059a882aa49adfeb162db17e5b72c118. Its host/voice.mjs
normalizes the live Codex thread/realtime/listVoices catalog and treats requested
versus confirmed voice/model as separate observations. It supplies a live
Codex/WebRTC voice integration, not bundled downloadable TTS weights or a generic
offline narration endpoint. No private code, pairing state, live sessions or
credentials were copied, and building this package does not depend on that repo.
The applicable design lesson is live discovery, explicit speaker selection and
honest model provenance. OpenAI voice names are not sent to NVIDIA's API.
Audio and rights
Initial cast: Magpie-Multilingual.EN-US.Jason.Neutral (Sam) and
Magpie-Multilingual.EN-US.Aria (Literate AI), observed from NVIDIA's live catalog.
Use the native sample rate and retain stock synthetic recordings with their
provider/model/voice, text/story/catalog hashes and audio identity. No real person
is cloned. NVIDIA's governing-terms notice
identifies the API Trial Service Terms. The package records rights context without inventing an
independent license grant for a model or a person's likeness.
Voice IDs record the requested selection from the observed catalog. The audio response does not separately report a confirmed effective voice or model-weight revision; neither is claimed. The named model is the API model family.
Plex fonts come from the pinned @fontsource/ibm-plex-sans package; its SIL Open
Font License is copied beside each generated companion. The Motion runtime's MIT
license and dependency licenses must remain with a distributed companion.
From audition to the full episode
- Preserve the user-approved audition voices and unchanged spoken dialogue. The performance score is implemented with locally estimated, recording-bound phrase cues. Render and technically verify the new cut, then hand it to the user for visual validation of communicative acting, listener reactions and readable holds. The old animation remains rejected; no visual approval is inferred for this revision.
- Expand into installation/acquisition, doctor, an observed adopt plan, explicit acknowledgement, and the first real verification/failure-recovery path. Capture fresh disposable-project evidence and bind the actual tested revision.
- Carry prerequisite and authentication explanations on screen. Mark walkthroughs, edited replays and unexecuted steps; never present generated terminal text as a successful execution transcript.
- Use the existing video CLI for final assembly/integrity, inspect every shot and listen to the complete soundtrack, then publish only to an authorized destination.
The existing accepted 11:52 film and its media remain intact. The native Windows qualification timeout is unrelated to this authoring pilot and is not resolved or relabelled by a successful animation or speech check.