Prompts — Fully Public

Every prompt actually used in Silent Era · The Guardian trailer production is here. Good prompts are my current production standard. Bad prompts are annotated with the problem, root cause, and real-world impact.

Prompt Architecture

The project's prompts are organized into five layers:

Global Rules → Negative Constraints → Style / Color Suffix
    ↓
Anchor Layer: Scene Anchors + Character Anchors + Object Anchors
    ↓
Shot Description Layer: Composition, Action, Mood, Camera
    ↓
I2V Motion Layer: Motion Templates + First-Frame Constraints
    ↓
Continuity Layer: Barrier Scale, Environment Continuity, Character Consistency

---

Good Prompts

1. Global Rules & Negative Constraints

These two blocks are the foundation for every prompt. Every line was burned in through rework.

Global Visual Suffix:

painterly cinematic realism, soft shadows, muted color palette,
16:9 aspect ratio, not raw photography, not cel-shaded, not game render

Global Negative Constraints:

no ancient palace, no fantasy costume, no hanfu, no xianxia, no game armor,
no glamorous hero design, no cyberpunk neon city, no cute anime chibi face,
no overly bright colors, no western medieval ruins, no mythological deity look,
no giant barrier dome, no huge shield over a city, no floating sky orb

These solved the early recurring problem: models kept interpreting "Chinese sci-fi" as xianxia, cyberpunk, or superhero aesthetics. Negative constraints proved more token-efficient and more stable than positive descriptions.

2. Barrier Scale Constraint

The most critical constraint in the entire project, and the single largest source of rework.

small human-scale spherical barrier, about 3 meters in diameter,
only large enough for three people,
near terrain or low-floating with terrain,
not a city-scale dome, not floating high in the sky, not covering buildings

Why this works: Early versions just wrote "3-meter barrier" and models completely ignored the number. Binding it to "three people + floor mat" as a proportional reference finally got models to render correctly. Every "not" clause exists because a model actually produced that error.

3. Anchor Generation Prompts

Anchors are the foundation of character consistency. Each anchor includes reference style, spatial relationships, and negative constraints.

Scene anchor examples:

ID	Prompt
SC02_BARRIER_INSIDE	Interior of a small 3-meter translucent pale-blue spherical barrier inside/overlapping an old apartment space, only enough room for three sleeping family members and Chen Mo, quiet blue membrane light, record book and scratch marks, cool grey shadows, minimal warm accents.
SC03_BARRIER_EXTERIOR	Wide modern concrete ruins under dark crimson sky, collapsed apartment blocks, exposed rebar, broken asphalt; a tiny 3-meter pale-blue spherical barrier sits near ground level like a small bead in the middle distance, buildings much larger than the barrier.

Character anchor examples:

ID	Prompt
CH01_CHEN_YOUNG	Chen Mo, 32-year-old ordinary Chinese warehouse worker father, short black hair, tired but gentle eyes, modern dark blue work uniform, average build, grounded working-class realism, not heroic.
CH06_XIAOYU	Xiao Yu, 2-year-old Chinese girl, pink bunny pajamas, round toddler face, natural child proportions, sleepy innocent expression, not anime cute, not older than 2.

4. I2V Motion Templates

All motion prompts derive from these templates:

Motion Type	Template
Blue Energy Forming	ethereal pale blue soul-energy [action], slow motion, soft bioluminescence, gentle particles, no explosion, no superhero effect, keep faces stable
Blue Energy Flowing	pale blue energy streams [flow path], gradual and organic, soft bioluminescence, subtle pulsing, no magic circle, no lightning blast
Character Micro-Motion	subtle human motion, natural restrained movement, [action], gentle breathing, not dramatic, keep character faces stable
Barrier Wave	translucent blue energy surface [action], gradual and organic, fragile membrane texture, small 3-meter human-scale spherical barrier, not shattering, not explosive

5. Color Mapping Table

Each scene has a mandatory color suffix to prevent tone drift across shots:

Scene	Mandatory Tone
SC01_HOME	warm tungsten interior, ochre and wood tones, soft practical lighting
SC02_BARRIER_INSIDE	pale blue bioluminescence, cool grey shadows, minimal warm accents
SC03_BARRIER_EXTERIOR	dark crimson overcast sky, collapsed apartment blocks, exposed rebar, broken asphalt, ordinary household debris, blue glow focal point
SC04_RUIN_SNOW	dark crimson overcast sky, grey-white snow over drifted modern city ruins, tiny low-floating blue cocoon focal point

---

Bad Prompts: 16 Lessons Learned

The following prompt strategies failed in production. Each is annotated with the problem, root cause, and real impact.

#	Pitfall	Impact
1	Not passing local image references — Agent assumed text descriptions = reference images	~20-30 wasted generations
2	Generating character-containing scenes before character anchors were locked	All early SC01-SC08 character candidates discarded
3	STYLE_C_WORLD reference inherited wrong barrier scale — model can't separate "style" from "content" references	3-5 wasted gens per scene
4	"3-meter barrier" text alone is unstable — models don't understand absolute dimensions	Highest-frequency rework cause in SC02 series
5	Three-color toy blocks became a visual nail — minor prop hijacked scene focus	Multiple SC03/SC08 candidates polluted with garish color symbols
6	"Cat blanket" in prompt caused models to generate cat-shaped objects	Spurious cat forms in SC01 home scene
7	SC02 tried to solve spatial structure AND emotional close-up in a single prompt	~27 SC02 candidates reworked
8	SC08 was consistently drawn as family interaction or city-scale energy field	~21 candidates reworked, second only to SC02
9	Prompt exceeded 5000-character API limit	Single generation rejected
10	Too many reference images for SC02 — model overfit to old images	Core cause of SC02 rework loops
11	API returned HTTP 200 with no image data	Agent blindly retried same prompt, wasting API calls
12	Dual-reference close-up alignment caused overfitting	Candidate Y output unstable
13	Re-running entire anchor scene groups repeatedly	Hundreds of wasteful generations across 5 scenes
14	Skipping the pre-production checklist	Early PU rework rates far higher than late-stage
15	PowerShell Chinese output garbled — terminal encoding not UTF-8	Agent parse failures and workflow interruptions
16	Default 5-second I2V is not a workflow — shot duration must be evaluated per-shot	Wasted initial 5s runs, re-renders needed

Patterns Across All 16 Failures

1. Skipping prerequisites — running before anchors were locked, skipping checklists
2. Treating "descriptions" as "constraints" — model doesn't obey "don't do X" without forced negative suffixes and proportional references
3. Fixing too much at once — re-running entire groups, multiple reference images, space + emotion simultaneously
4. Agent blind retries — same prompt fails, Agent retries unchanged, wasting tokens and API calls

---

Recommended Adoption Order

1. Copy the global rules + negative constraints as your every-prompt foundation
2. Generate your character/scene/object anchors before any character-containing shots
3. Assemble each shot's prompt from the template structure — don't hand-write full sentences
4. Derive I2V from motion templates, binding quantitative constraints (seconds, resolution)
5. Build your own negative constraint list — add one entry for every new problem you encounter