What Belongs in Scene Notes vs Narration vs Style Lock
Stop mixing prompt concerns: scene notes hold must-show visuals, narration holds spoken lines, and style lock holds look rules that never change mid-scene.
What Belongs in Scene Notes vs Narration vs Style Lock
Stop mixing prompt concerns. Scene notes hold must-show visuals, narration holds spoken lines, and style lock holds look rules that never change mid-scene. Teams that dump everything into one mega-prompt regenerate for the wrong reason.
Storyboard-first AI video fails less often when each beat separates what appears, what is said, and how it looks. That split is a brief template, not a tool feature.
The three-bucket template
| Bucket | Owns | Never owns |
|---|---|---|
| Scene notes | Subjects, props, on-screen nouns, camera intent, must-avoid clutter | Voice tone, brand essay, model names |
| Narration | Exact VO (or on-screen text if mute-first), timing cues | Wardrobe, palette, lens |
| Style lock | Palette, line weight, character still refs, lighting, do/don’t | Per-scene plot twists |
Print this as three columns on every scene card. If a sentence could live in two buckets, rewrite it until it cannot.
Why mixing concerns burns credits
A typical failure looks like: “Friendly host in soft light explains pricing while teal flat UI cards float, cinematic, 4k, no text.” That one string asks the model to solve plot, VO, and style at once. When the face drifts, people “fix style”; when the VO is wrong, people “fix the whole scene.”
Separate buckets make QA local:
- Wrong object → edit scene notes only
- Wrong words → edit narration only
- Wrong look → edit style lock (or re-approve stills), then re-check all scenes
This is the operational version of storyboard-first vs generate-then-fix.
Scene notes: write nouns a stranger can draw
Good scene notes answer: If audio is off, what must still teach?
Include:
- Subject (one primary)
- Supporting props (≤3)
- On-screen labels (exact strings if any)
- Framing (close / medium / wide) and safe zone for 9:16
- Must-avoid (extra hands, fake logos, busy dashboards)
Exclude:
- Adjectives that are really style (“cinematic,” “dreamy”)
- Long explanations better said in VO
- Multiple teach promises in one beat
Marketers briefing without a designer can use the same card discipline—see how marketers brief an AI storyboard without a designer.
Narration: one job per spoken line
Narration is the teach line for that beat—not a voiceover essay.
Rules that keep VO honest:
- One claim per scene
- Same nouns as on-screen text (voiceover vs on-screen text)
- Word count that fits the beat length (roughly 2.5–3 words/sec spoken)
- No style instructions inside quotes
If the line needs a second idea, split the scene.
Style lock: approve once, reuse everywhere
Style lock is a short, boring document:
- Character still package (face, wardrobe, props)
- Background family
- Palette + contrast rules
- Forbidden: real logos, readable micro-UI, extra characters
Approve stills against this lock before any motion (character still consistency). Do not renegotiate palette in scene 4 because “this beat feels more premium.”
Filled SaaS example: 4-beat “shared annotation” Reel
Style lock (shared): flat vector, navy/teal cards, one host silhouette from approved still, no photoreal faces, no tiny UI chrome.
| Beat | Scene notes | Narration |
|---|---|---|
| 1 Hook | Split frame: messy rewind arrow vs pin icon; text “Pin the moment” | “CSMs: stop rewatching the whole call.” |
| 2 Problem | Timeline scrub with highlight chaos; no faces | “Context dies when notes live in three chats.” |
| 3 Mechanism | Pin marker + share card with teammate avatars as dots | “Pin the moment, share the clip.” |
| 4 CTA | Simple card: “Try on your next call” + product noun | “Annotate once. Brief the team.” |
Notice style never appears in the VO column. Notice narration never lists hex colors.
QA checklist before you generate
- Every scene has three filled buckets
- No style adjectives inside scene notes
- Narration nouns match on-screen nouns
- Style lock referenced by ID/name, not re-described each scene
- One teach promise per beat (45-second scene checklist)
- Aspect ratio already chosen (9:16 vs 16:9 vs 1:1)
Prompt paste patterns that still respect buckets
When a tool forces a single text field, concatenate with hard labels—do not merge meanings:
STYLE_LOCK: <id or short paragraph>
SCENE_NOTES: <nouns + framing + must-avoid>
NARRATION: "<exact line>"If the model ignores labels, shorten each bucket rather than adding more adjectives. Long prompts that restate style five ways usually fight the still reference.
Team RACI for the three buckets
| Bucket | Drafts | Approves | Forbidden editors |
|---|---|---|---|
| Scene notes | PMM / creator | Creative lead | Legal rewriting visuals into prose |
| Narration | PMM / sales enablement | Claims owner | Design changing meaning via synonyms |
| Style lock | Design / brand | Brand owner | Anyone mid-batch “for fun” |
Write names, not roles only. Anonymous RACI is how drift returns.
Common contamination examples (fix these)
- Scene notes: “cinematic host passionately explains…” → move cinematic to style; passionately to VO direction if needed; keep host + prop nouns
- Narration: “in our beautiful teal world…” → delete brand poetry; keep teach nouns
- Style lock: “also show the new pricing table from Q3…” → that is scene content, not style
Run a red-pen pass looking only for contamination before you approve stills.
Soft tool note
Whatever stack you use, cleaner approvals happen when each beat separates visual / VO / style before render. Platforms such as Scro are built around approving that storyboard gate; the three-bucket template works even if you paste cards into a spreadsheet.