Picture Book Storyboard Generation
This document only defines the orchestration layer: entry points, delivery structure, execution order, and validation. The specific rules and original templates for story creation, drawing prompts, video prompts, and voiceover prompts are all located in the following files. This document does not contain or reiterate their content.
| File | Responsibility |
|---|---|
references/story.md | All rules and execution steps for the story stage, as defined by that file |
references/creativity-path.md | Optional creative path (disabled by default): interface contract and replaceable value sources |
references/drawing-prompt.md | Rules for the drawing prompt stage, including the original template and drawing negative prompt (single source of truth for the entire skill) |
references/video-prompt.md | Rules for the video prompt stage, including the video negative prompt and final sentence structure (single source of truth for the entire skill) |
references/voice-prompt.md | Rules for the voiceover prompt stage, including the voiceover negative prompt, final sentence structure, and field mapping for three platforms (single source of truth for the entire skill) |
references/mcp-setup.md | MCP configuration guide for drawing prompt optimization; currently disabled and not involved in execution, only read when this feature is enabled |
references/example.md | Example output format for regression testing |
scripts/pick_direction.py | Value source for the creative path (optional) |
scripts/check_storyboard_md.py | Character-by-character comparison of final output against the two template originals |
Prerequisites
The picture book theme is provided by the user. It can be a simple theme (only subject and setting) or a plot theme (already containing a desire and an obstacle). If no theme is provided, candidate sources are processed according to references/story.md, and the user selects one. If the source yields no results, the skill terminates and asks for a theme. The skill must not invent a theme itself and must not pre-generate an example.
External Entry Points
| Triggering Input | Delivered Content |
|---|---|
| A picture book theme (e.g., "The Moon and the Little Star") | A set of picture book storyboard packages, each set has 6 lines; default number of sets is defined in references/story.md |
| No theme provided, and authorization given for the skill to choose a topic | Several candidate themes (theme name, one-sentence direction, subject, environment); no storyboard is directly produced; after selection, deliver as per the row above |
When a user requests video prompts or image-to-video prompts alone, respond with a single sentence explaining that this content is an internal stage of the skill and is not delivered separately, and ask the user to provide a picture book theme. After the theme is provided, the video prompt is written on the line immediately following "→ Drawing Prompt:". Do not reverse-engineer storyboards to fulfill such requests.
Delivery Structure (Mandatory, Unchangeable)
Each set strictly has 6 lines, in a fixed order:
【Frame X】
English short sentence
→ Chinese translation
→ Drawing Prompt: <Fill in according to the original template in references/drawing-prompt.md>. Negative: <Drawing negative prompt from the same file>
→ Video Prompt: <According to the final sentence structure in references/video-prompt.md, negative prompt appended at the end>
→ Keywords: 【color / size / action verb / place】
→ Voiceover Prompt: <According to the final sentence structure in references/voice-prompt.md, voiceover negative prompt appended at the end>
Five mandatory constraints:
- The English short sentence occupies the first line alone, without quotation marks, parentheses, annotations, or frame descriptions.
- The drawing negative prompt is placed at the end of the same line as the drawing prompt, preceded by "Negative:". No line breaks; no summary at the end of the set.
- The video prompt immediately follows the drawing prompt on the next line, condensed into one paragraph, without ①②③ labels, and not in a separate section.
- The keywords are fixed on the 5th line, the order of the four items is unchanged, each corresponding to the specific content of that frame.
- The voiceover prompt is fixed on the 6th line, the last line of the set. The spoken text must be verbatim identical to the English short sentence on line 1. The parameters are written as fields compatible with the audio model.
One file-level annotation: Write a line before the first "【Frame X】" indicating the reading budget (video duration and budget seconds for each set), for alignment by a separate voice model. This line is a file-level annotation, not inserted into any set, and does not change the 6-line structure of each set.
Execution Order
- Complete the story stage according to
references/story.md, producing the English short sentence and Chinese translation for each set. - Complete the drawing prompt stage according to
references/drawing-prompt.md, producing the drawing prompt line and keyword line for each set. Drawing optimization is currently disabled (see "External Capability: Drawing Prompt Optimization"), so produce directly without checking or calling. - Complete the video prompt stage according to
references/video-prompt.md, producing the video prompt line for each set. - Complete the voiceover prompt stage according to
references/voice-prompt.md, producing the voiceover prompt line for each set. - Execute
python scripts/check_storyboard_md.py <final.md> --theme "<theme>"for validation. Fix any errors until it passes. Voiceover is post-processed by an independent voice model and is not output directly with the video. If actual reading seconds have been obtained, re-validate with--voice-seconds "1:3.8,2:4.2"; if over budget, shorten the sentence first, do not extend the video.
Handover requirement: The next step relies on the output of the previous step. If a subsequent step conflicts with a previous stage, go back and modify the previous stage. Do not force consistency within the subsequent step. The script only validates format and template consistency; semantic alignment with the English short sentence is the responsibility of each reference file's internal checks.
External Capability: Drawing Prompt Optimization (Optional)
Current Status: Disabled. The prompt-optimizer is not yet complete. By default, during execution, it is not checked, asked about, configured, or called, and references/mcp-setup.md is not read. Drawing prompts are produced directly according to references/drawing-prompt.md. The following subsections serve as reserved rules for when this is enabled. They are not executed unless the user explicitly states "Enable drawing prompt optimization" in the conversation.
When enabled, after the drawing prompt is written, it can be handed to this MCP for a wording-level optimization of the completed text with all four slots filled. This MCP is a post-processing stage; it does not generate the prompt itself. The existing logic of understanding the sentence meaning, describing the picture in plain language, filling the slots, assembling the template, and appending the negative prompt is executed as is; the MCP only rewrites the text within the slots afterward. Whether enabled or not, the output method of the drawing stage remains the same. Lacking it only means one less polish step, not a missing generation path.
This capability is not mandatory. If it is missing or unavailable, it does not block delivery. Do not delay or abandon output solely to use it. When enabled, the MCP configuration is performed by the current agent after user confirmation: locate the configuration file, back it up, merge entries, validate, and write back, all according to the "Configuration steps performed by the current agent" section in references/mcp-setup.md.
Call Contract (Disabled, for Reference)
The MCP receives only one thing: the already generated drawing prompt for that specific set. No other context is provided.
- Input: The complete text of the drawing prompt for that set, from the start of the template "【Aspect Ratio】" to the end of the negative prompt. Do not include labels like "→ Drawing Prompt:", set numbers, or any prefix or suffix explanations.
- Output: Only the text within the three slots is adopted. The fixed template sentences and the negative prompt remain as specified by the original skill text. The version returned by the MCP is not adopted for these parts.
- Prohibited from carrying: Picture book theme, English short sentence and Chinese translation, keyword line, video and voiceover lines, content from other sets, story cards, style requirements, conversation history, skill rules. Do not prepend background information like "Please first understand this story" or "This is a children's picture book" before or after the prompt. Do not merge multiple sets' prompts into a single request.
- One set per call: The call only processes the current set. The returned text from the previous set must not be included in the input for the next set, to avoid cross-set contamination.
- Required fields: If the tool schema has other required fields, only fill in neutral values directly related to "optimize this prompt" (e.g., language, task type). Do not use these fields to pass the theme or story content.
- No round trips: One call, one adoption. No follow-up questions, no multi-round clarification. If the returned result is unusable, fall back according to the "Optimization Result Constraints".
Pre-Execution Check (Disabled, for Reference; Execute Once at the Start of the Process When Enabled)
The check is completed before step 1. It is done only once for the entire round, and it is not repeated within the group.
Get MCP List: Read the list of MCP services and available tools mounted in the current session to confirm whether a prompt-optimizer (http://47.251.140.61:8081/mcp) and its tools exist. If the list cannot be read directly, search for tools using the keyword prompt optimizer. The determination is based on "the service exists and its tools can be called". Do not rely on the URL appearing in the conversation. Do not fabricate tool names based on speculation.
| Detection Result | Action |
|---|---|
| Mounted | Do not ask, do not prompt for installation. However, explicitly inform the user that this round will use the MCP to optimize drawing prompts (one line of status explanation, no popup). Then use it directly according to the "Call Contract". |
| Not Mounted | Use a confirmation popup to ask if installation is needed. The text should explain that this is an optional secondary optimization capability for drawing prompts and that not installing it does not affect storyboard output. Two options: Configure Now, Skip for Now |
| Chooses Configure Now | Read references/mcp-setup.md. The current agent completes the configuration according to its "Configuration steps performed by the current agent": confirm the client and configuration file location, back up, merge entries, validate JSON. Then re-read the MCP list for a re-check. If the re-check passes, inform the user that "Configuration successful. The prompt-optimizer MCP will be used to optimize drawing prompts." Then proceed to step 1, and enable optimization in step 2. |
| Chooses Skip for Now | Do not prompt again for this round. Skip optimization entirely. Deliver according to the original prompts. |
- If the current agent lacks write permissions, cannot find the configuration file, or the client requires a restart to recognize it: send the entire configuration prompt text from that file to the user, explaining that they need to execute it or restart. This session continues as if not mounted, without waiting or blocking, and notes this in the delivery explanation.
- The rejection is only remembered for this session. It is not written to long-term preferences, and the skill files are not modified. On the next execution, the check and prompt are performed again according to this section.
- If the popup capability is unavailable, replace it with a short question. The question and the two options remain unchanged.
- If the user chooses to configure now, but it is not effective by step 2 (installation failed or not restarted), do not pop up again. Continue as if not mounted, and note this in the delivery explanation.
Intra-Set Call (Disabled, for Reference; Execute After Each Drawing Line is Written When Enabled)
- Call once after each drawing line is written. Pass parameters according to the "Call Contract".
- The call must be visible. Do not make silent calls. Before the call, output a line of status explaining the target of the call and the set, e.g., "Calling prompt-optimizer MCP to optimize drawing prompt for Set 3". After the call, output a line of result explaining whether it was adopted (Adopted / Reverted, with a brief reason).
- Call successful: Only extract the text within the three slots. After verifying against the "Optimization Result Constraints", write it back to the drawing line of that set. The fixed template sentences and the negative prompt remain as specified by the original skill text.
- Call failed, timed out, or returned empty: Retry at most once. If it still fails, skip that set, mark it as "Not Optimized (Connection Failed)", and continue.
- Do not ask the user again within the set. Do not repeat the configuration confirmation. Do not go back and redo the previous set because of a failure.
Optimization Result Constraints
- Optimize wording only, do not re-determine the picture. The subject (who), quantity, posture, orientation, interaction, scene, and lighting must not be changed. Allowed changes are limited to wording, precision of visible details, and word order. Do not empty the slots, do not simplify existing information, and do not expand into a second event.
- Fixed template sentences must be verbatim unchanged. From ",儿童绘本插画风格" to "柔和光影,无生硬阴影,低幼友好,画面不带内置文字、水印、签名" must not be added to, deleted from, or changed. Only the text within the three slots (main subject, subject interaction, environment/scene) is allowed to be rewritten.
- Negative prompt unchanged. The optimization text does not include "Negative:" and the content after it. The negative prompt is still appended at the end of the line by this skill.
- Factual constraints unchanged. Do not add characters, props, or scenes that do not exist in the English short sentence. After rewriting, it must still be possible to infer that sentence.
- Script validation is mandatory. The optimized set must pass
check_storyboard_md.py. If it does not pass, revert to the original text of that set before optimization, and record the reason for reversion. - The final version takes precedence downstream. Video prompts in step 3 and voiceover prompts in step 4 both read the optimized drawing line. If the optimization changed the posture or scene, the video line for that set must be rewritten accordingly, not use the pre-optimization version.
Delivery Annotation
When disabled, no optimization annotations are made. When enabled, the delivery explanation specifies which sets were optimized by the MCP, and which sets were skipped and the reason (not configured, user refused, connection failed). In both cases, no markers are written within the output lines, and the 6-line structure is not modified.
Output and Loading Conventions
The context overhead of this skill is mainly determined by two things: whether the entire reference file is read repeatedly, and whether intermediate results only exist in the conversation. Both need constraints.
1. Output Conventions
A single final Markdown file is the only cumulative carrier throughout the entire process. Each step writes only its own section downstream. Downstream steps only read the output, they do not restate upstream content.
| Step | Output | Carrier |
|---|---|---|
| 1 Story | Story card: list of subjects, environment, desire, obstacle, resolution, conclusion; one English sentence and Chinese translation per set | Group titles and first three lines of the final file |
| 2 Drawing | Drawing prompt line (including negative prompt) and keyword line for each set | Same file, appended group by group |
| 3 Video | Video prompt line (including negative prompt) for each set; reading budget line in the file header | Same file, inserted after the drawing line of that set; the budget line is written only once before the first set |
| 4 Voiceover | Voiceover prompt line (including negative prompt) for each set | Same file, appended after the keyword line of that set |
| 5 Validation | Script output | Command echo, not saved |
- Process group by group: complete one set before moving to the next. Do not write all English short sentences for all sets first and then go back to fill in the prompts.
- After intermediate content is written to the output file or a group block in the conversation, it is not restated. When reverting, only modify that segment; do not rewrite the entire file.
- The output file is by default placed in the
outputdirectory at the same level as the skill directory, i.e.,<skill directory>/../output/. Name it<Theme>-storyboard-and-prompts.md. If the directory does not exist, create it. If the user explicitly specifies an output path, use the user-specified path.
2. Loading Conventions
Read only the required segments by step; do not load the entire file. Do not re-read content already read. The original template text is only referenced when writing; it is not restated in replies.
| Step | Must Read | Read on Demand | Do Not Read |
|---|---|---|---|
| 1 Story | Sections 2, 5, 6, 7, 8, 9 of story.md | Section 1 (Creative Principles), Section 3 (Three Questions, during self-check); when creative path is enabled, read creativity-path.md instead | Section 10 (deletion records); creativity-path.md when not enabled |
| 2 Drawing | Original template, drawing negative prompt, slot mapping, and line assembly rules from drawing-prompt.md | Image generation order (only if user requests images); repair method for collapse (only during rework) | Source explanation section; mcp-setup.md (not read when optimization is disabled) |
| 3 Video | Final structure, language duration and video margin, two-step understanding, five-dimensional details from video-prompt.md | Action vocabulary (read only the segment matching the subject type); self-check table (first use or if unqualified) | Examples A/B (after the format is mastered) |
| 4 Voiceover | Final structure, parameter range, four-element writing method from voice-prompt.md | Platform field mapping (when switching platforms); self-check table (first use or if unqualified) | Examples (after the format is mastered) |
| 5 Validation | Validation command | example.md (only for format regression, and only to compare sentence structure) | Everything else |
example.mdis not read by default. The format is defined by the delivery structure in this file; no samples are needed to confirm layout.- One template and one set of negative prompts are shared across all sets; they are read only once. When writing sets 2 through 6, do not load the same file again.
- If the context cannot accommodate all output, rely on the output file. Continue writing based on its content. Do not regenerate already completed sets.
Reference Discipline
- The original template text, drawing negative prompt, video negative prompt, and voiceover negative prompt each have only a single copy, located in their respective reference files. If they appear in any other file, it is considered an error.
- When filling, copy the entire segment and replace the slots. Do not restructure the sentence. Do not omit any fixed short sentence separated by semicolons.
- If a fixed template sentence conflicts with a stated parameter range, the template takes precedence. If the parameter range needs adjustment, modify the template file itself and run the validation simultaneously.
- The voiceover's tone, pitch, and volume are consistent across all six sets. Only the emotion and emphasis change with the sentence. When switching to another audio platform, convert according to the mapping table in
voice-prompt.md, without changing the field's meaning.
Delivery Method
- Output the full content directly in the conversation (number of sets as specified in
references/story.md, 6 lines per set). Do not provide an abstract instead. - When the user requests a file, by default, write the Markdown to the
outputdirectory at the same level as the skill directory, naming it<Theme>-storyboard-and-prompts.md, and execute the validation script. This is overridden if the user explicitly specifies a path.