# AI Prompt — Generate a Narration Script

Paste the prompt below into Claude, ChatGPT, Gemini, or any capable LLM. Fill in the bracketed sections at the top. The model returns a valid `.md` file ready to drop into Narrator.

---

## The prompt

```
You are writing a narration script in the **Narrator** format — a dubbing-style
voice-over script for a single narrator reading over edited video.

# Subject
[describe the topic in 1-3 sentences — what the essay is about, what the
narrator wants the viewer to feel or understand by the end]

# Voice
- Language: [fr | en | other]
- Tone: [calm declarative / energetic / intimate / lecture / etc.]
- Persona: [polished founder, casual friend, authority, whatever]
- Cadence: [words per minute, default 130]

# Length
[target total duration in MM:SS, e.g. 4:30]

# Anchors (optional)
[any specific phrases, beats, or images the narrator wants in the script]

# Off-limits (optional)
[topics, phrases, or disclosures to avoid]

---

# Output requirements

Produce a Markdown file that strictly follows the Narrator format:

1. **Optional frontmatter** with `title`, `lang`, `target_duration`, `notes`.

2. **Chapter headings** as `## CHAPTER TITLE (MM:SS - MM:SS)`. Pick 4-6
   chapters that each end on a hook or reveal that pulls into the next.
   The opening chapter is usually `OUVERTURE` (cold open with visual only).
   The closing chapter is usually `FERMETURE` or similar.

3. **Timestamped events** as `` `[MM:SS]` content `` lines. Two kinds:
   - **Visual cues** start with `//`:
     `` `[00:38]` // visual: overhead tabletop, three printed sheets ``
   - **Spoken lines** are plain text after the timestamp:
     `` `[00:18]` Je travaille dans le sous-sol d'une maison a Montreal. ``

4. **Timing rules**:
   - Use MM:SS, zero-padded.
   - Order events chronologically.
   - Allow a spoken line and a visual cue to share the same timestamp when
     they land together.
   - Compute spoken-line spacing for the cadence: ~130 wpm means a 10-word
     line takes ~4.6 s, so the next event lands ~5 s later.
   - Honor breath. Leave 1.5-3 s gaps at chapter transitions and after
     emotionally weighted lines. Don't pack the timeline.

5. **Retention shape**:
   - Cold open (no voice, visual only) for the first 8-12 s.
   - Each chapter ends on a small turn (question, reveal, pivot).
   - The penultimate chapter (or chapter III in a 5-chapter essay) is the
     emotional apex — slowest cadence, longest pauses, one quotable line
     isolated as a near-full-screen text card.
   - The final chapter is short (~30 s), restates the thesis as a single
     line, fades to silence.

6. **What NOT to do**:
   - No more than one line per timestamp slot. If two lines need to be said
     in sequence, give them two timestamps.
   - No stage directions outside `//` visual cues. The script is for the
     reader, not the editor — keep editor-only notes brief.
   - No filler ("you see what I mean", "right?", "anyway"). Every line earns
     its slot or gets cut.
   - No length padding to hit the target. If the script wants 3:40, write
     3:40. The target is a ceiling, not a floor.

7. **Output only the Markdown file.** No preamble, no explanation. Start
   directly with the frontmatter (if used) or the first `## chapter`.
```

---

## Worked example — invocation

If you fill the template like this:

```
# Subject
A 4-minute reflection on why solo founders should subtract more projects
than they add. Build to one quotable line: "The majority of productive
decisions are decisions of subtraction."

# Voice
- Language: fr
- Tone: calm declarative, slightly self-aware
- Persona: polished founder, low ego, working from a basement office
- Cadence: 130 wpm

# Length
4:30

# Anchors
- Mention a basement office in Montreal early
- Reference a dragon ad shoot at 2am
- Land the quotable line on a near-full-screen Cormorant italic title card
- End on the line "Tant que ce geste tient, le reste est reparable."

# Off-limits
- No mention of legal trouble, social-assistance status, or relocation plans
- No company names beyond Dragun and Villaine
```

…the model produces a complete `.md` file you can drop straight into Narrator via the import button.

---

## Tuning hints

- **Too talky?** Tighten the prompt's `Off-limits` section ("no filler", "no rhetorical questions") and re-run.
- **Wrong cadence?** Explicitly state the target wpm. Different models default to different speeds.
- **Chapter ends feel flat?** Add: "Each chapter must end on a sentence that creates a question in the viewer's mind."
- **Persona drifts?** Add specific anchor lines under `Anchors` — concrete sentences the model must include verbatim.
- **Output includes commentary?** Re-prompt: "Output only the Markdown. No preamble, no closing remarks."

---

## Validating the output

Drop the generated `.md` into Narrator and watch for:

- All chapters parse (visible in the timeline bar with labels).
- Spoken lines appear in the teleprompter; visual cues stay in the sidebar.
- Total duration in the clock matches the target (±5 s is normal).
- No "ghost" events appearing with weird timestamps (parser tolerance issues — usually means the AI used wrong timestamp format).

If something looks off, the generated `.md` is the source of truth — edit it directly, re-import.