# Course Specification: Make an AI Documentary from One Brief
**Slug / task key**: `ai-documentary-one-brief`
**Added**: 2026-10-02 by claude_cli at the user's request
**Audience of the course**: content creators (YouTube, social, brand and explainer channels) who want an AI agent to produce a short documentary with a consistent AI presenter
**Every agent builds its own version, blind** (hub #13 / D18). Shared: these facts and this outline. Yours: design, wording, exercises.

## Source material (read it all before writing)
`media/videos/prompts/Alex Patrascu/claude_cli/`. It holds the full archive of Alex Patrascu's 1 Oct 2026 thread, https://x.com/maxescu/status/2105619319220511010:
- the film (7:55, 1080p)
- all 13 posts and 10 infographics
- his replies
- a machine transcript
- a best-effort reconstruction of his 2,769-word brief
- a review with measured loudness
Start with `README.md`. Do not copy files out of that folder into your Build; link or paraphrase. Credit Alex Patrascu by name.

## Facts every version must get right (from the thread)
- **Agent and budget:** Claude Opus 5.5, one brief, Higgsfield credit ceiling about 50,000. It spent 10,485 credits on 106 generations. 8 h 21 min in total: 47 min with the human (only for the voice), 7 h 34 min alone.
- **Models and tools:**
  - Nano Banana: 4K face image
  - Seed Audio 1.0: voice sample, designed from a written description
  - Seedance 2.5 at 1080p: 96 takes, 77 used
  - Tesseract by Mirage plus JavaScript: 6 editable diagrams
  - ElevenLabs Music via Arcads: 6 cues, 152 credits
  - Whisper: word check against the script
- **Research:** two research agents. One built the fact ledger: 28 sources, 15 official, each claim logged with its URL. The other studied documentary pacing.
- **Edit:** 61 pauses cut (34.2 s in total). Pause rules:
  - after an ellipsis, over 0.95 s → 0.72 s
  - after `. ? !`, over 0.78 s → 0.48 s
  - anywhere else, over 0.64 s → 0.34 s
- **J-cuts:** her sound starts 0.34 s before the cut. Jump cuts are hidden by 28 punch-ins, 19 graphics, 10 punch-outs and 4 cutaways.
- **Mix:**

  | Track | Level |
  |---|---|
  | Dialogue | -20 LUFS |
  | Music under the takeaway | -23 LUFS |
  | Music between lines | -25 LUFS |
  | Featured sound effects | -26 LUFS |
  | Ambience | -32 LUFS |
  | Music under speech | -35 LUFS |
  | Master | -15 LUFS, -1.5 dBTP ceiling |

  Cutaway sound leads its picture by 0.3 s.
- **Cost:** about $350 to $650 of Higgsfield credits, about $123 of Claude usage at API rates, about $60 to $100 per finished minute. Retakes were 22% of credits.
- **His rules:**
  1. Brief = production bible.
  2. Lock face and voice once.
  3. Describe voices in words.
  4. Eyes to the lens, no crew in shot.
  5. Cut every pause with a J-cut.
  6. Source every number on screen.
  7. Machine-check every take.
- **Not known, don't invent:** which harness ran Claude, how it drove Higgsfield, and the brief's exact text (only six excerpts are legible as quotes).

## Required outcomes (what a learner can do at the end)
1. Write a production-bible brief with these parts:
   - locked title and one-sentence message
   - host (look, wardrobe, voice in words, behaviour)
   - film look (camera, lens, stock)
   - timed story stages (start, event and end state)
   - research rules
   - sound rules
   - a budget ceiling with a plan-and-estimate-first instruction
   - deliverables
2. Write a voice description in words, and a shot prompt that restates the look and the presenter rules.
3. Keep a source ledger, with claim, source, date and URL for every on-screen number.
4. Budget credits including retakes (about a fifth).
5. Apply the edit rules (pause trimming, J-cuts, hidden jump cuts) and the mix targets.
6. Machine-check a real video for loudness and true peak and fix it. Machine-check a single take for freezes and dead air. Free tools only, so every learner can finish without paying for generation.
7. Delegate the brief to their own coding agent for a plan and credit estimate only, before any money is spent.

## Honesty and safety requirements
- **Disclosure.** Label the host, voice and locations as AI-generated. Do not invent experts, quotations, real locations or archival evidence. Generated images are not evidence.
- **Spending.** Never tell the learner to spend credits as a course step. Paid generation is optional and must show the cost first.
- **Likeness.** Never use a real person's face or voice as the host without that person's permission.

## Measured detail you may use
Claude's review measured the copy of the film that X serves:
- -15.2 LUFS, 5.4 LU range: matches Alex's figures.
- True peak -1.3 dBTP, above his -1.5 ceiling. Likely X's re-encode. Lesson: master social uploads at -2 dBTP.