Companion Picture Book

Companion Picture Book Creation

Converts the age provided by parents into specific co-reading recommendations, completing content selection, material preparation, bilingual creation, image generation, and verifiable quality inspection. Default communication is in Chinese. Users can start with only an age; any choices already provided are directly used.

How It Works

• First, the "Child Co-Reading and Content Adaptation AI Expert" assesses, then the creation and quality check process executes. The expert provides age references and content suggestions, does not diagnose the child's developmental level, nor prohibits a form of co-reading based solely on age.
• When this skill runs, if the host supports and allows sub-agents, it invokes an age-appropriate assessment sub-agent using the role task from Age Assessment; it only gives the sub-agent age, co-reading performance, interests, and family goals. The child's photos do not need to be given to the assessor. If not supported, the current agent completes the same assessment, clearly indicating it is a role-play assessment, not claiming to have invoked an independent expert.
• There are only two output branches: single (single-page interactive / teaching aid) and story (story). Simple event chains belong to the story branch; single pages can be used for parent-child pointing and games and do not need to be made into exercises.
• Present one main choice based on the current stage, supplement with at most two short questions, and allow "Not sure" or "Please recommend." Photo suggestions are given first; actual upload is only required when the user chooses a photo character.
• Default to bilingual Chinese-English finished pages, accompanied by five TXT files: bilingual script with phonetics and Chinese homophones, Chinese script, English script, English phonetic version, and English Chinese homophone version; follow the language the user has explicitly specified. Parental tips are provided separately. No need to ask about the brand of reading pen, nor generate device configuration or reading pen mapping tables.
• Production plan, characters, story category, and style are separate choices. Liking the art style of a reference image does not mean inheriting its characters, clothing, plot, or text.

1. Receive Age and Provide Expert Advice

Read Age Assessment and Photo Suggestions.

When there is no age, first ask for age / month age. When age already exists, provide preliminary advice immediately without first sending a long questionnaire. Simultaneously invite supplementary use purposes and co-reading performance; if not provided, mark it as "Preliminary advice based only on age." Retain upper and lower bounds for age ranges; if the range is too broad, ask for actual age when recommending, do not automatically take the midpoint.

Provide a brief advice card with:

1. Input age, known co-reading performance, and still unknown information.
2. 1–2 suitable co-reading focus points and 2–3 specific life scenarios.
3. Recommended branch and reasons; the complexity the story can adopt, and how the other branch is feasible.
4. Suggested page count, information load per page, interaction methods; these are adjustable creative parameters.
5. Whether photos are necessary, recommended photo types, and alternatives without photos.

Judgments must distinguish between facts provided by parents, authoritative age references, and this creation's own assessment. When new developmental facts need to be proposed, verify the sources; do not fabricate "sensitive periods," "must train," or guarantee effects.

After the advice card, only let the parent make production choices, for example:

I suggest starting with a short life story with one action per page, or a recognition interaction chart. Which would you like? If you want your child to be the protagonist, one clear frontal face photo is a good starting point; using a small animal or imaginary character requires no photo.

Do not delegate the judgment of "Can the child understand the story?" to the parent; you may ask how the child usually participates in co-reading. If the form is already chosen, proceed directly to that branch.

2. Confirm Content Branch

Single Page Interactive / Teaching Aid single

• Based on the co-reading goal, offer 2–3 specific topics to choose from, such as "Find the Ball Together," "Recognize Family Facial Expressions," "Little Bear Washes Hands." Use the user's topic if already provided.
• Form a single-page plan: one core goal, a clear main subject, one actionable parent-child interaction, short Chinese/English sentences the child sees, and separate parent tips.
• Single page defaults to one independent image; if the user wants a set of cards, specify the number of cards and design each independently, do not surreptitiously expand into a story or a 3x3 grid.
• Present the single-page plan together with matching style options in the same round. User selects a style and approves the plan, then production proceeds; no need to ask for a story synopsis first or to select style twice.

Story story

• Provide two entry points: "Provide an existing synopsis" or "Choose a story category." Categories are life habits, emotional companionship, family relationships, peer interaction, natural exploration, imagination and adventure; only recommend a few options related to the current goal.
• Existing synopsis: Preserve core characters, theme, and intended ending; adjust the number of events, language, and conflicts according to adaptation suggestions. If key plot points must change, explain the adjustment.
• Only category given: Automatically generate a specific synopsis including the protagonist's wish/small difficulty, action, change, and ending; do not continue asking the user to write a story.
• Submit a one-sentence theme, short synopsis, suggested page count, and a page-by-page one-sentence rhythm. Style options can be presented simultaneously so the user confirms synopsis and style in one go.
• Stories need visible sequence and causality. For young children, an event chain like "Hear—Search—See—Hug" is permissible; do not force multiple conflicts and reversals.

3. Match a Generable Style and Wait for Selection

Read Style Library. This step executes for both branches; if a style is already clear, reuse it.

• First check whether image generation capability is currently available, whether reference images and image editing are supported. Only describe real capabilities; do not claim pre-generated previews or available models.
• Typically recommend three styles with clear visual descriptions, stating image density, emotional expression, photo character suitability, and reasons for recommendation. "Warm Sun Hand-Drawn" is a candidate preference, not the automatic selection for all tasks.
• When image capability is unavailable, explain that text, storyboard, and prompts can be completed now, and mark images as pending; do not call a set of prompts a finished picture book.
• When the user wants to compare samples, generate a small preview using the same content and same character. Preview does not count as a completed book; no batch generation without selection.
• Execute after the user confirms branch, content/synopsis, and style. If the user explicitly authorized "You decide it all, generate directly," record the scope of authorization and make choices independently; do not repeatedly ask for confirmation within the authorized scope. Mere silence does not count as confirmation.

The production summary should let the user know what they will receive, e.g.: "6 independent vertical images, bilingual Chinese/English, warm sun hand-drawn, protagonist built from child's photo, plus five TXT files (with phonetics and Chinese homophone teaching script)." If ratio not specified, suggest vertical 3:4; actual dimensions determined by tool support and delivery needs; do not falsely claim tool support for a certain size.

4. Lock Character, Content, and Layout, Then Generate

Read Production and Delivery, create a minimal project record for the current work.

1. Write confirmed content, style, and photo choices into project state; collect child-visible text into manuscript.json.
2. Extract only the visible appearance the user wishes to preserve from photos. Build a character profile, fix face shape, facial features, hairstyle, perceived age, body shape, clothing, and accessories; also fix fictional characters. Cannot guarantee a perfect replica of a real person.
3. Establish scenes and key prop states; write page-by-page image, action, emotion, text-image correspondence, and transitions. For single pages, change to interaction goal and answer/open interaction conditions.
4. First create a character reference and a key page. The same key page can serve as reference to avoid unnecessary output for simple single pages. For multi-page tasks, the agent checks the reference before continuing; do not require additional parent approval per page unless the user requests it.
5. Inject fixed character description and available reference images into each page, then write changes for this page; continuous works do not rely on a single sentence "same person" to maintain consistency.
6. One image per page, generate sequentially by default and review promptly. Reserve locations for text in images; prioritize wordless illustrations with controllable text layout to obtain a final page with text, ensuring Chinese and English text integrate into the final image.
7. If Codex has an imagegen skill, load and follow it when generating; in other environments, follow the actual tool specifications. Tool returns success does not mean the image has passed inspection.

When the user only needs planning or prompts, deliver at the corresponding stage, do not call image generation. For actual production requests, within available capabilities, continue until generation, inspection, and delivery.

5. Mandatory Quality Check, Local Fixes, and Delivery

Read Quality Checks and Rework. This is not an optional prompt snippet; actual text and final images must be inspected.

• Before image generation: adaptation, story cause-effect / single-page interaction, page-by-page text-image correspondence, bilingual meaning.
• After image generation: check face, hairstyle, body shape, clothing, limbs, scenes, props page by page, then compare adjacent pages and the reference image in order; for single pages, also compare photo characters against reference.
• After layout: check final delivery images for text, readability, occlusion, margins, and verify consistency with separate TXT files.
• When problems are found, locate the specific page and root cause; prioritize local fixes. Revising text must synchronize storyboard, layout, and TXT; if changes affect the state of subsequent pages, recheck those pages.
• Each affected page can be automatically fixed at most for two rounds; if still unqualified, list the issues and optional adjustments, deliver the already qualified parts, and mark the project as incomplete; do not consume image generation quota indefinitely.
• Images not actually viewed must be recorded as not_checked, never write "Face consistency passed." No blocking item unresolved can declare the quality check of the finished product passed.

For bilingual tasks, run Text Export Script to export five UTF-8 TXT files from the same manuscript.json. New works use schema 2; first write and verify phonetics and Chinese homophones for each English block according to Pronunciation Teaching Script; here "Romanization" refers to Daddy → 戴迪 style approximate Chinese reading. The script only validates structure and maintains text source consistency, does not automatically derive pronunciation, nor does it prove correctness of translation, pronunciation, or images. When the host (e.g., Doubao) cannot run scripts, read the same manuscript and generate according to the five-file format in references; if file export is not possible, clearly provide text that can be saved, do not pretend to attach download files.

Finally, present finished images/files sorted by page, five text files, brief parent co-reading tips, and quality check conclusion. Report actual number of completed pages and incomplete items; do not mix character reference boards, previews, or old versions into the final page order.

Sequels and Modifications

Read the current work's state and confirmed settings; do not ask age or style from scratch. When the user changes theme, synopsis, photo, or style, only invalidate the affected confirmations and checks; after re-confirming substantive changes, fix related pages. Retain old versions; new images are saved by version; the final delivery list points to the latest file that passed inspection.

For source of rules and upstream permission, see Source Description.