All Things Grow Prompt Structure Snapshot 20260712

# All Things Grow Full-Canvas Structure Snapshot 2026-07-12

This is a dated evidence boundary for a read-only audit of the Neowow canvas
`https://neowow.cn/video-works-view?id=147`. The user supplied the logged-in
Chrome session. The assistant used the page's own authenticated workflow
client to export the complete canvas graph without copying or cloning the
project and without exposing the login token.

## Evidence Level

- `observed-full-canvas`: 362 nodes and 533 edges were read from the project
  JSON on 2026-07-12.
- Node mix: 180 image, 148 video, 23 audio, and 11 group nodes.
- Prompt-bearing nodes: 111 video prompts and 81 image prompts.
- Status: 105 of the prompt-bearing video nodes were marked `SUCCESS`, but
  only one video generation node explicitly recorded `reviewed=true`.
- Asset inventory inside node data: 1,186 URL occurrences resolving to 507
  unique media URLs: 173 images, 323 videos, and 11 audio files.
- Prompt reference resolution: 824 `@[asset-id]` occurrences, 116 unique IDs,
  and 108 IDs resolved exactly to a unique media URL by basename. Eight stale,
  placeholder, or externally removed IDs did not resolve.
- `inferred`: repeated structures support reusable production hypotheses.
- Not causal: no same-model A/B run proves that length, a named style, or any
  single clause caused quality or reach.
- `SUCCESS` is a provider job state, not a creative acceptance label.

## Density Is Task-Typed

The canvas does not support the rule "successful prompts are always long."
Among the 105 successful prompt-bearing video nodes:

- 16 prompts were under 200 characters;
- 19 were under 1,000 characters;
- 59 were at least 2,500 characters;
- median length was 2,719 characters;
- maximum length was 7,404 characters.

The 111 video prompts contain five useful operational cases, but keyword-based
classification counts are heuristic rather than ground truth:

| Family | What owns the complexity |
| --- | --- |
| director-dense generation | the video prompt owns timed execution |
| compact generation | one visual event or experiment |
| storyboard-delegated generation | a grid/storyboard owns sequence detail |
| dialogue/audio cue | line, voice, delivery, and timing own the task |
| surgical video edit | an existing video owns everything except one delta |

This means prompt density must follow control ownership. Do not expand a
storyboard delegation, voice cue, or one-delta edit into a director-dense
rewrite. Do not compress a prompt-owned multi-beat scene into a mood summary.

## Upstream Asset Factory

The 81 image prompts were often denser than the video prompts:

- median length: 3,209 characters;
- 75th percentile: 10,224 characters;
- maximum: 18,003 characters.

A separate visual audit of all 180 image nodes, including uploads, captures,
generated assets, and edit variants, assigned these primary duties:

- 40 character assets;
- 36 key/start/end frames;
- 29 edit/repair/variant assets;
- 19 palette/color-script cards;
- 17 prop/weapon assets;
- 14 storyboard/staging boards;
- 9 environment assets;
- 8 poster/layout outputs;
- 7 creature/vehicle/mecha assets;
- 1 near-black unclassified image with no reliable downstream evidence.

The asset layer carried substantial production truth before video generation.
Observed asset duties included:

- character face/body/costume identity;
- environment geography, set dressing, and material family;
- prop or weapon shape and state;
- first frame, end frame, and cross-clip handoff;
- storyboard order and representation-layer changes;
- color-card palette with named HEX duties;
- action cadence or camera-motion reference;
- voice identity, dialogue delivery, and beat timing;
- existing-video inheritance for surgical edits.

Across all 533 graph edges, 439 originated at image nodes, 52 at video nodes,
and 42 at audio nodes. The 23 audio nodes resolved to 11 unique audio tokens,
showing deliberate reuse and duplicate placement. Asset IDs were not
decorative: prompts usually named which shot, character, prop, palette, frame,
cadence, or voice each reference controlled.

## Repeated Director Controls

Counts below are lexical evidence from all 111 video prompts, not quality
scores:

- 102 used explicit reference anchors;
- 93 specified camera geometry;
- 93 specified sound, ambience, BGM, white noise, or effects;
- 88 described local micro-actions;
- 80 described foreground/middle/background or screen placement;
- 77 encoded a causal trigger or reaction chain;
- 75 used timing or start/middle/end action stages;
- 74 used negative or exclusion locks;
- 67 controlled dialogue or voice performance;
- 53 used semantic color, color cards, or HEX values;
- 49 used hard cuts, whip pans, match cuts, masks, or transitions;
- 48 used must-show/must-not-show or off-frame rules;
- 42 declared continuous-take or physical-continuity rules;
- 35 used an explicit style-lock section.

The strongest repeated pattern was not adjective density. It was a causal
execution chain: