Repository documentation

Literate AI course creative brief

This is this repository's story, not inherited framework policy. Derived projects choose their own presenters, tone, format, visuals and subject. The generic author-instructional-videos skill teaches the production craft; this document owns LitAI and Sam.

Creative direction

Current revision (1.2): settle the argument, before CI

The ten-minute "earn a small trial" cut spent too long persuading Sam that narrow wins justify adoption. The 1.2 film inverts that. LitAI opens with a compelling, concrete picture of agentic software development done properly: reviewed promises, independent acceptance, three-platform builds, an installed package that ran and a commit-bound receipt, all before CI. Sam reacts badly to "the usual AI slop sales pitch". He is the ringleader of the holdouts in an argument raging at his company (and across the industry): the CTO wants agent-first everything, and Thursday's architecture review decides.

LitAI then takes Sam's beliefs one at a time and settles each with evidence. Every debunk doubles as the tutorial step that proves it, so the audience learns installation, planning, generation, independent acceptance, bounded repair, pre-CI release gates, HTML views, brownfield adoption, learning and releases as a side effect of the argument. The seven beliefs are: agent code is unreviewable; the agent grades its own homework; CI is the safety net; an agent is a black box; adoption means a rewrite; agents never learn; and releases are where it breaks. The ending settles the argument both ways: agents do not ship on vibes, and hand-reading every generated line was never the safety net. Evidence is.

Sam is not a straw man. He keeps real objections and the last laugh; LitAI still concedes real limits at the moment they matter (a lenient oracle passes, run-history controls are not shipped, remote steps are walkthroughs, later adoption stages are not shown complete). "Superhuman" means stamina, not judgment: the harness runs the whole gate every time. The CI chapter names conventional CI/CD pitfalls (push and pray, queue latency, discovering platform failures last, untested pipeline files holding production keys) and answers them with make release-check and litai release qualify (this host, then workers, then CI only where needed).

Speech text and display text are separate. Pronounce AI as individual letters, likewise CLI and SSH; say C plus plus. Preserve literal commands in captions and visuals. New expressive local voices require a fresh audition: acceptance of the previous Kokoro opening does not approve either these voices or this script.

LitAI is a capable, warm female instructor with dry wit. Sam is a competent male C++/Make engineer whose habits leave packaging and release proof under-specified. Neither is a lecture delivery device. Sam makes choices, tries commands, notices results, objects, and occasionally gets the last laugh. LitAI responds to those actions instead of continuing a prewritten monologue. Make is not the villain: unrecorded assumptions and unproved shipping decisions are the problem.

Use short exchanges and earned callbacks, not alternating paragraphs. The running joke is Sam's final-final-really-final.zip, not his intelligence. Humor should make a boundary memorable, then get out of the way. No generic corporate music bed is required; clean voices and intentional pauses matter more.

Visual language

Use the two original character portraits at readable scale with shot/reaction changes; avoid postage-stamp mascots over a bullet deck. Show a binary moving from a developer's laptop toward another machine, the specification-to-product path, a real terminal, and the evidence or package produced. Reveal only the relevant boundary at the moment it is discussed. Commands remain legible and copyable from companion documentation. Do not put the entire spoken explanation on the screen. Captions occupy a reserved lower band.

Greenfield narrative

1. Cold open: Sam has a compiled program and calls it shipped. LitAI asks what actually reached the user. Show the gap, rather than listing lifecycle stages.
2. Acquisition: show github.com/jordanhubbard/literate-ai, the README and installation documentation. Distinguish current development checkout from a published release; do not invent a PyPI package or nonexistent release asset.
3. Setup: demonstrate one complete macOS/Linux checkout route: Python 3.11+, Git and Make; git clone, cd, and the documented make install path. Explain prompts for prerequisites, the reported installation prefix/PATH, and verify litai version and litai doctor. Link the documented Windows alternative.
4. Agent readiness: choose and authenticate a supported coding CLI separately; never display credentials or claim installation alone grants model access.
5. Create: Sam reviews the read-only onboarding plan, then acknowledges apply. Show the actual files and explain the hello specification by changing one observable requirement, not by describing the entire taxonomy.
6. Code and proof: capture a real coding session and actual tests/acceptance; keep meaningful failures and fixes, trim waiting with an on-screen time-edit label. No invented terminal output or animation labeled as screen capture.
7. Packaging and shipping: show the actual wheel and verification. A local wheel is not a published release. End with an actionable next step and the ZIP-name callback, not another lifecycle summary slide.

Production gate

The project owner accepted the neural audition's direction and requested one continuous video play. Begin with character introductions and their perspectives; make the real coding/adoption work the middle, and close with Sam reluctantly convinced but asking about updates and linked repositories. Use chapter navigation, not disconnected snippets. Humor targets unverified shortcuts and industrial-speed AI slop, never the audience's intelligence. Public feedback publication is authorized. The continuous production package owns the script and fresh evidence.

The original three slide-based cuts are rejected creative drafts, not the target quality. First produce a short replacement opening with locally generated neural voices, back-and-forth, visual action, and readable subtitles. Keep originals for comparison. Obtain reviewer feedback on that cut before spending on a full-series rerender. Record which voice/model and visuals were actually reviewed; technical media verification alone does not constitute creative acceptance.