ig-carousel-writer

IG carousel post

First decide what the reader should understand, then let the assets do the work of explaining. By default, take the instructional direction of "real-world assets + clear labeling," interspersing concise text pages depending on the subject matter. Save the deliverables back to the official location specified by the user or in the current project. First read references/寫作與事實規則.md。

Direct sharing by the author (one fish, many dishes)

When adapting the author's own video, spoken words, or article to another platform, the speaker is still the author. Readers should be able to understand this content without needing to see the source.

• The main text should directly express opinions, suggestions, and well-grounded experiences. Do not introduce your own content with phrases like "I mentioned in the video," "the approach in the video," "the author shared in the video," or "this post is compiled from the video," and do not merely switch to "I once said" to retain a reported-speech framing.
• For example, "I like to use a breakdown method I shared in the video" can be changed to "I think the best method is…"; "the video mentioned that you can have regular gatherings" can be changed to "My suggestion is to schedule regular gatherings." Naturally choose "I think," "My approach is," "I would suggest," or state it directly according to the meaning; do not force "I" into every sentence.
• Changing the tone cannot create facts: only write "my approach" when there is a basis in actual practice; suggestions are still suggestions; do not turn other people's experiences into your own, and do not fabricate long-term habits, results, original attribution, or guarantees out of thin air. "I think it is best" can express the author's position, but does not automatically expand into an objective best.
• Books, research, and other people's methods still retain necessary attribution or citation. The author's own source URLs, transcripts, timecodes, and adaptation records go in internal source/review notes; the public main text should be independently complete. If there is a genuine traffic-driving purpose, one "watch the full video" CTA may be retained at the end.
• Before delivery, check paragraph by paragraph for "in the video / in the film / in the clip / once mentioned / compiled from the video" and sentences that introduce the author in the third person; rewrite according to context, do not mechanically delete words. Original transcripts, explicit quotations, analysis of others' content, and internal editing timecodes are not affected by this rewriting rule.

1. Read the source and brand

• Read the most recent AGENTS.md, the specified main draft, and existing brand documents; do not ask again for information already provided. If the source is published content, verify the actual published version; do not pass off a local draft as the published version.
• Confirm the topic, audience, and the one thing readers can do after finishing. Copy should only use claims that can be confirmed in the main draft; photos cannot endorse cases, results, or experiences that have no source.
• Read references/品牌設定模板.md. If brand settings already exist, follow them. If colors are missing, you may continue with a provisional direction of white background and near-black text; do not fabricate unknown accounts, avatars, or logos, and do not force in placeholder avatars.
• When the user only wants copy/planning, deliver the text and asset list only; do not generate images or create HTML. Planning can be completed without images; list missing items in the shooting checklist.

2. Decide the assets and visual logic

First read references/real-world-assets.md and references/design-system.md. When a visual reference is specified, first look at the original image provided by the user; learn only the reading flow and information hierarchy, and do not appropriate assets, attribution, or frameworks.

1. Write the core message of this set in one sentence, and choose an object, situation, or interface familiar to readers to explain it.
2. Divide materials into: publishable originals, research-only, pending confirmation, and generated mockups. Publicly visible does not equal reusable; a realistic appearance does not equal an actual photo.
3. On each page, write down the image's function: explain a relationship, provide evidence, locate an operation, build a metaphor, compare differences, or showcase a resource. If you cannot answer what its function is, delete the decorative image.
4. When comparing the same thing, keep the composition, cropping, and label positions consistent; when explaining steps, use the same scene and add annotations step by step. Do not change styles on every page just for the sake of variation.
5. Use real screenshots, book covers, or photos only when the source and the image support it; if you cannot find them, list specific self-shot/image-acquisition needs, and do not fabricate a successful download.

3. Write page-by-page copy and storyboard

According to references/output-template.md save the copy, material ledger, and production specs. Default to 5–10 pages; when the content requires it and the platform's current limit allows, you may add more, and do not treat the source's 16-page case as the platform spec. If the main draft is too short, make a shorter carousel; do not pad it.

• The opening can use a specific question, a direct contrast, an object metaphor, or a single claim; it does not have to always be "pain point → punchline → CTA."
• One key point per page. A title of about 8–20 Chinese characters and body text of about 30–80 characters is an editorial starting point; do not sacrifice necessary context for word count; if overloaded, split the page first.
• On material-led pages, let readers first recognize the object/interface, then read the annotations and conclusion; on text-only pages, build hierarchy with title, explanation, and examples.
• You do not need to put an avatar, Tweet Card, border, or punchline box on every page; essential text must not be cropped. Use whitespace for grouping and pauses.
• Traditional Chinese, natural Taiwanese phrasing; do not appropriate others' first-person stories, proprietary frameworks, free-resource promises, or comment/DM funnels. When citing someone else's method, give attribution.
• The caption supplements the usage context, necessary sources, and one suitable action. Do not force line breaks sentence by sentence, and do not impose Emoji, hashtag counts, or follow requests; the CTA should link only to resources that actually exist and can be delivered.

4. Produce images

• Default final spec is 1080 × 1350 (4:5); if the user specifies a ratio, that takes priority. Make a representative page first, then extend it to the whole set, completing them in order.
• When you need to generate/edit raster images, load imagegen; if the environment does not have the imagegen skill, use the currently available and authorized image tools; according to the reference-image method supported by the tool, first view the local original image, then pass in the actual image; do not merely mention the filename in the prompt. Clearly indicate which one is the edit target, inserted material, or style reference.
• The real content of original photos, screenshots, and logos is a preservation item. For interface text, numbers, and logos that are required to remain exactly unchanged, do not promise that generation tools can preserve them pixel by pixel; first generate an independent visual layer, keep the original in an editable layout layer, and use available layout tools. When other raster editing methods are needed, handle them according to the current tool rules and user authorization.
• Do not force every original material to be regenerated; do not call generated images actual photos, and do not call interface mockups product screenshots. Before sending private materials externally, confirm authorization.
• Use the original logo files specified by the author and that they have the right to use. When using generation/editing tools, actually pass them in; do not redraw or switch versions; if the file is unreadable, report the missing item. Do not preload instructor logos or personal avatars.
• Do not default to an HTML screenshot deliverable. Only go the HTML route when the user explicitly wants HTML or an editable webpage; the output requirements of the content workflow take priority. When tools are missing, deliver usable planning, missing items, and the unfinished scope; do not call a storyboard a completed PNG.
• Proofread Traditional Chinese text precisely, character by character; fix typos first, then review again. Preserve the original draft, materials, prompts/edit instructions, and selected output.

5. Visual and delivery acceptance

Open the output one by one, and review the rhythm with a full contact sheet; when only planning, check only the copy/storyboard and do not claim the visual final product has passed.

• In a phone preview about 360px wide, the title, main explanation, and necessary screenshot information are readable; if the details that need to be read are too small, crop a local section and preserve the context.
• Necessary text does not go out of bounds, images are not distorted, and materials are not obscured by overlapping masks; arrows have clear start and end points, and labels do not point to the wrong objects.
• The same color and symbols have consistent semantics across the whole set; do not rely on color alone, pair them with text.
• The original interface numbers, people, products, logos, sources, and content claims read back consistently. Research images are not mixed into the publishable folder.
• The purpose of the materials on each page is clear; pages that need real materials do have the originals, or are marked as pending in the plan, and are not padded with generic AI backgrounds.
• The copy and caption on each page are directly shared by the author, without a tone of relaying one's own sources; internal materials/production notes are not mixed into public content.
• The caption has no fabricated CTA; file names, page numbers, ratios, order, versions, and missing items have been checked.

Finally, provide the output location, actually completed items, and pending items. Attach a representative image/contact sheet for easy viewing; if not published, do not claim it has been published.