Myreels Storyboard

Step 1: Collect User Requirements

Collect the following information:

After collecting inputs:

• Use the current OpenClaw session's default LLM to understand the story semantically
• Infer core roles, recurring relationships, and phase-by-phase visual beats from the user's actual plot, not only from surface keywords
• Do not ask the user to configure an extra LLM just to use this skill
• Do not rely on a bundled local storyboard script for core planning; OpenClaw should draft the CSVs directly
• The old helper scripts/generate_storyboard.py is deprecated and should not be referenced in this workflow

Step 2: Character Design (Important!)

Before shot design, establish character consistency using a three-view character sheet (三视图).

Default rule:

• Treat character design as a single-image turnaround sheet, not as a poster
• The first-pass deliverable should show the same character's front view, side view, and back view on one canvas
• Keep outfit, body shape, hairstyle, and accessories identical across all three views
• Use full-body neutral standing pose and a plain background unless the user asks otherwise
• Recommend nano-banana2 first for generating the base three-view character sheet
• Prefer a character-first workflow: approve the character sheet and tags first, then generate storyboard shots with those tags injected

Character Design Fields

Three-View Diagram Output

Generate character descriptions for AI image generation. The output should be structured for a single image containing all three views:

CHARACTER: [Name]
STORY_ROLE: [protagonist / opponent / companion / love_interest]
RELATION_TO_MAIN: [self / conflict / ally / romance]
TAGS: [Consistency tags for prompt injection]
REFERENCE_MODEL: [Usually nano-banana2 for the first three-view sheet]
SINGLE_IMAGE_TURNAROUND_PROMPT: [English prompt explicitly requesting one image with front / side / back full-body views]
NEGATIVE_PROMPT: [English suppression terms for layout drift]

FRONT: A [age] [ethnicity] [gender] with [hair], [eyes], [build], wearing [outfit], neutral standing pose
SIDE: Same character from side profile showing [distinctive feature / silhouette / accessory placement]
BACK: Same character from behind showing [hairstyle back shape / outfit back detail / asymmetry]

Prompt rule:

• The single_image_turnaround_prompt must front-load the layout requirement: one image, same character, front view, side view, back view, aligned left to right
• Do not output three unrelated portrait prompts as the default character design result
• Only add expression sheets, weapons, or alternate outfits after the base three-view sheet is approved

Character Consistency

For each character, build a character tag library that will be reused in all shot prompts:

# Example character tags
"短发女孩, 左脸有痣, 红色耳钉, 黑色皮夹克"
→ "short hair, beauty mark on left cheek, red earrings, black leather jacket"

Character-First Mode

Recommended operating mode:

• Generate characters.csv first
• Generate relationships.csv for recurring character pairs
• Lock the main character's single_image_turnaround_prompt, negative_prompt, and character_tags
• Only define extra character rows when the story actually implies them; do not invent an opponent, love_interest, or companion just because the genre suggests one
• Reuse the approved character_tags in every visual_prompt
• Treat the first approved character row as the anchor for downstream storyboard drafts
• In multi-character stories, inject only the characters relevant to each shot phase instead of dumping everyone into every prompt
• Build relationships.csv as a pairwise matrix for all recurring character pairs, not only protagonist-centered pairs
• Reuse relationship_type and prompt_hint when designing confrontation, ally, or romance shots
• Also use visual_relation, emotional_relation, and blocking_pattern to keep staging and mood consistent across repeated pair shots
• Let relationship_type bias the default shot language: conflict leans toward low-angle / confrontational framing, ally toward cooperative full shots, romance toward closer eye-level coverage
• Let theme bias the baseline shot rhythm too: action leans kinetic, romance leans intimate, thriller leans uneasy, sci-fi leans scale/discovery
• In scenes with 3+ recurring characters, choose the active pair/group per shot based on the current beat; do not force every established relationship into every shot

Default OpenClaw behavior:

• The skill should use the active OpenClaw default LLM to draft characters.csv, relationships.csv, and storyboard.csv
• The CSVs should reflect narrative understanding, not just mechanical keyword expansion
• Keep CSV column names stable in English, but make user-facing cell values follow the user's language whenever possible
• Keep AI-facing prompt fields in English for cross-model portability: single_image_turnaround_prompt, negative_prompt, front_view_prompt, side_view_prompt, back_view_prompt, character_tags, prompt_hint, and visual_prompt
• Keep IDs and machine-oriented codes stable: character_id, relationship_id, shot_id, scene_code
• visual_prompt should preserve concrete story specifics from the user's outline in English, including names, props, locations, and beat-specific actions
• description and action should vary shot by shot; avoid repetitive generic phase labels when a more specific beat description is available
• For mixed-language input, keep description in the dominant user language; if no dominant language is clear, follow the user's most recent instruction language
• If a downstream automation flow later needs normalized English enum values, create a secondary English export instead of replacing the user-facing review CSV

Step 3: Generate Storyboard

Shot Structure

Each shot includes these fields:

Shot Type Reference

Camera Angles & Psychology

Camera Movement

• Fixed / Static
• Dolly In (推近) / Dolly Out (拉远)
• Pan (摇镜) / Tilt (倾斜)
• Follow / Tracking (跟拍)
• Orbit (环绕) / Crane (升降)
• Handheld (手持) / Breath-like (呼吸感)

Shot Difficulty Grading

Emotion Mapping

Each shot must specify target emotion:

Step 4: AI Prompt Engineering

Prompt Structure

[Subject] + [Action] + [Environment] + [Lighting] + [Quality] + [Style] + [Camera]

Prompt design rule for this skill:

• single_image_turnaround_prompt should use layout-first language for nano-banana2
• visual_prompt should carry theme + phase cues so action / romance / thriller / sci-fi drafts do not collapse into the same generic shot language

Required Quality Tags

Always include:

• cinematic
• 8k or high resolution
• shallow depth of field or bokeh
• dynamic lighting

Control Tags

If needed, add:

• no text / no watermark / no distorted face
• solo / two-shot / crowd

Character Consistency in Prompts

Always inject character tags from Step 2:

# Before
"A woman fighting aliens"

# After (with character tags)
"A young woman with short hair, beauty mark on left cheek, red leather jacket fighting aliens, cinematic, 8k..."

Scene Consistency

Use scene code +固化环境描述:

SCENE-A: "abandoned colony base, red dust, two moons, ruins"
# All shots in SCENE-A reference this environment

Step 5: Output Confirmation

Multi-Platform Output

Feishu: Use feishu_bitable for online editing Telegram: Formatted text table or CSV attachment Other: Default to CSV

Storyboard Table Fields

shot_id, scene_code, shot_type, camera_angle, movement, duration, description, visual_prompt, action, dialogue, emotion, sound_fx, notes, difficulty, image_url, video_url, status

Note:

• The CSV headers stay in English for schema stability
• Human-facing values in all three CSVs should follow the user's language
• The visual_prompt field is always in English for AI image generation
• If automation needs English enums such as shot_type, camera_angle, movement, emotion, difficulty, or status, write a separate English export copy instead of mutating the review CSV

Step 5A: Project Workspace Management

When this skill writes files, it should automatically create a project workspace in the user's current working directory or requested output directory.

Default root pattern:

storyboard-projects/YYYYMMDD-project-slug/

Recommended structure:

storyboard-projects/YYYYMMDD-project-slug/
├── 00-brief/
│   ├── story-brief.md
│   └── user-notes.md
├── 01-characters/
│   ├── drafts/
│   │   ├── characters.v1.draft.csv
│   │   └── character-design.v1.draft.md
│   ├── approved/
│   │   ├── characters.approved.csv
│   │   └── character-design.approved.md
│   ├── exports/
│   │   └── characters.program-en.csv
│   └── generated-images/
│       ├── raw/
│       ├── approved/
│       ├── requests/
│       ├── tasks/
│       └── metadata/
├── 02-relationships/
│   ├── drafts/
│   │   └── relationships.v1.draft.csv
│   ├── approved/
│       └── relationships.approved.csv
│   └── exports/
│       └── relationships.program-en.csv
├── 03-storyboard/
│   ├── drafts/
│   │   ├── storyboard.v1.draft.csv
│   │   └── storyboard-notes.v1.draft.md
│   ├── approved/
│       ├── storyboard.approved.csv
│       └── storyboard-notes.approved.md
│   └── exports/
│       └── storyboard.program-en.csv
├── 04-generation/
│   ├── storyboard-images/
│   │   ├── raw/
│   │   └── approved/
│   │   ├── requests/
│   │   ├── tasks/
│   │   └── metadata/
│   └── videos/
│       ├── raw/
│       ├── approved/
│       ├── requests/
│       ├── tasks/
│       └── metadata/
├── 05-review/
│   └── review-log.md
└── 06-delivery/
    └── manifest.md

Workspace rules:

• Do not write user project artifacts back into the skill definition folder itself
• Create the project root before saving the first CSV, markdown note, image, or video
• Save editable drafts under drafts/
• After user confirmation, freeze a copy under the matching approved/ folder instead of overwriting history
• Save character three-view turnaround images under 01-characters/generated-images/raw/
• Move or copy user-confirmed character reference images into 01-characters/generated-images/approved/
• Save storyboard shot images under 04-generation/storyboard-images/raw/
• Save storyboard videos under 04-generation/videos/raw/
• Move or copy user-confirmed storyboard media into the matching approved/ media folder
• For every myreels-api generation, save the exact request payload under requests/
• Save task lookup snapshots and task IDs under tasks/
• Save response metadata under metadata/, including model name, mode, aspect ratio, seed if available, source image references, created time, and returned URLs
• Keep request / task / metadata filenames aligned with the output artifact name so reruns can reuse the same context
• Record major review decisions in 05-review/review-log.md
• Summarize the final approved artifact set in 06-delivery/manifest.md

Versioning rules:

• Use file names like characters.v1.draft.csv, storyboard.v2.draft.csv, images-shot-S01-01.v1.png
• Use *.approved.* for frozen user-confirmed files
• When revising an approved artifact, create a new draft version instead of mutating the approved snapshot
• Put normalized program-facing CSV exports in exports/ with names such as storyboard.program-en.csv and relationships.program-en.csv
• Use aligned sidecar records such as character-C001-turnaround.v1.request.json, character-C001-turnaround.v1.task.json, character-C001-turnaround.v1.meta.json
• Use the same pattern for storyboard shots and videos, for example shot-S01-01.v2.request.json and shot-S01-01.v2.meta.json
• If the user already provided a target directory, follow it and only apply this structure inside that location

Re-run stability rules:

• Before re-running a generation, read the latest matching request, task snapshot, and metadata files first
• Reuse the last approved prompt, model, aspect ratio, and source asset references unless the user explicitly asks to change them
• If the rerun changes any material parameter such as model, prompt, aspect ratio, seed, duration, or source image, write a new versioned request file instead of overwriting the old one
• In review-log.md, note why the rerun happened and what changed relative to the previous attempt

Step 6: Guide to myreels-api

Guidance Message (Image Generation)

Storyboard confirmed. Now use `myreels-api` skill for image generation.

Recommended workflow:
1. Generate one single-image character turnaround sheet first (front/side/back on one canvas)
2. Prefer `nano-banana2` for the base three-view character sheet, then generate key storyboard shots using: nano-banana2 / seedream 5.0 / kling v3 image
3. Verify character consistency across all shots
4. Confirm images before video generation

Say: "Use myreels-api to generate storyboard images"

Guidance Message (Video Generation)

Images confirmed. Now use `myreels-api` skill for video.

Recommended models (image to video):
- Kling O3 / Kling V3 (general video)
- Seedance 1.5 Pro SE (high quality)
- Wan 2.6 i2v / Hailuo-2.3 (quick preview)

Note: For complex shots (🔴), consider longer duration or multiple takes.

Say: "Use myreels-api to convert these images to video"

Step 7: Quality Control Checklist

For each shot, verify:

• Character matches design (outfit, hair, facial features)
• Lighting direction consistent with scene light source
• Motion trajectory physically possible (no clipping, floating)
• Shot type serves narrative rhythm
• Safe zone reserved (vertical: top/bottom 20% no key elements for subtitles)
• Emotion matches intended audience feeling

Output Files

• characters.csv - Character design data
• relationships.csv - Relationship matrix for recurring characters
• storyboard.csv - Complete storyboard
• assets/characters-template.csv - Character CSV starter template
• assets/relationships-template.csv - Relationship CSV starter template
• assets/storyboard-template.csv - Storyboard CSV starter template
• Character reference images (via myreels-api)
• Storyboard image URLs (via myreels-api)
• Project review notes and approved snapshots under the project workspace
• Request payloads, task snapshots, and response metadata for reproducible reruns

Notes

• visual_prompt must be in English with quality tags
• User can input in Chinese / Japanese / English
• In the three CSVs, user-facing values should follow the user's language; only AI-facing prompt fields stay in English
• If the workflow needs machine-friendly English enums, generate a separate *.program-en.csv export instead of changing the review CSV
• Each shot: 2-5 seconds recommended
• Always establish character design before shot breakdown
• Use difficulty grading to manage AI generation expectations
• In OpenClaw, this skill should primarily rely on the active session model for story understanding and CSV drafting
• Do not require users to provide extra LLM credentials just to use this skill in OpenClaw
• No local storyboard-generation script is required; draft the CSVs natively in OpenClaw
• Automatically create a project workspace so drafts, approvals, images, videos, and markdown notes remain organized
• Persist generation requests and metadata so image/video reruns can stay stable
• See references/ for detailed guides