Facebook Diy Video Workflow

Facebook DIY Video Workflow

Use the bundled original prompts without rewriting them.

Preserve Prompt Integrity

• Do not edit, rewrite, optimize, translate, condense, extend, or merge any bundled prompt.
• Read the selected reference file completely before using it.
• Keep the stored prompt text unchanged. Supply product details and attachments as separate user inputs.
• Do not fill an unfinished field or infer a missing production step unless the user explicitly provides that information.
• Pass new constraints alongside the original prompt instead of modifying the prompt file.

Select a Prompt

• For a natural lifestyle human-model first frame, read and use references/human-model-prompts/natural-lifestyle-first-frame-prompt.md verbatim.
• For the detailed Chinese beauty-presenter first frame, read and use references/human-model-prompts/beauty-presenter-first-frame-prompt.md verbatim.
• When a human-model first frame needs controlled product scale, natural two-handed contact, multiple independent versions, or product-center mosaic refinement, read and apply references/human-model-product-hold-size-refinement.md after selecting the source prompt. Keep this operational supplement separate from the immutable prompt.
• For a Seedance request about detailed dot-drill, diamond-painting, pearl-painting, numbered-circle placement, fuse beads, Perler beads, or equivalent precision bead craft, read references/seedance-detailed-bead-placement-prompt.md completely and preserve it as the immutable field-tested example. Apply the detailed bead-craft route below before video generation.
• For any other Seedance request, read and use references/seedance-fast-10s-prompt.md verbatim.
• For a parent-child DIY request, read and use references/parent-child-diy-image-prompt.md verbatim, then continue through the Veo/Grok prompt as defined below.
• For Veo, Veo 3, Veo 3.1, Grok, Grok prompt, multi-scene video switching, or continuous multi-angle camera movement, read and use references/veo-grok-multi-scene-prompt.md verbatim.
• For a Facebook mature-audience spoken sales script or the user's "GPT full workflow," read and use references/gpt-full-workflow.md verbatim.
• For DIY reference-video analysis, production-step extraction, and a 15-second AI-video recreation script, read and use references/video-breakdown-recreation-prompt.md verbatim.
• After generating a spoken script, use references/multi-scene-product-display-prompt.md verbatim by default.
• If the input, script, or requested workflow mentions unboxing or opening the package, skip the default multi-scene branch and use references/factory-old-man-unboxing-prompt.md verbatim.
• If the user explicitly requests both workflows, run them separately in the requested order. Do not combine their source text into a third prompt.

Screen Complex Products

Before any image or video generation, inspect the product for complex wires, cords, branches, endpoints, connectors, repeated clips, dense edges, grids, transparent overlaps, or an exact functional topology.

1. When any high-risk signal is present, read and apply references/complex-product-generation-risks.md.
2. Recommend avoiding topology-critical generation, full disassembly, full assembly, or exact routing shots by default. For company work, include a concise recommendation to discuss a shot conversion with the manager or project lead before spending generation credits.
3. Do not treat a successful storyboard image or grid as proof that image-to-video is feasible. State that temporal consistency can still require many rerolls even when every still looks correct.
4. Produce the required feasibility note before generation. Preserve the communication purpose of the shot while proposing a finished-result view, presenter shot, voice-over, real close-up, or another simpler visual.
5. Continue only when the user still requires generation. Validate the first generated image against the exact structure at full size, then restrict each generated shot to one local area, one connector, or one simple action.
6. Stop the image-to-video branch if a critical part, path, endpoint, count, connection, or perspective relationship is wrong or drifts during motion. Use the reference's real-footage, static-image, or editing fallback instead of attempting to hide the error with more camera movement.

For detailed bead-craft work, do not reject the shot only because it contains repeated beads or a grid. Route it through the dedicated first-frame gate, restrict the shot to one verified macro area and one-piece-at-a-time placement, and stop when the grid, numbered position, completed state, piece count, or contact relationship is wrong.

This preflight supplements the selected original prompt without modifying it.

Route Human-Model First Frames

Apply this route when the user mentions a human-model prompt, real-person first frame, talking-head character image, presenter master image, or equivalent wording in any language.

1. Treat this as an image stage that runs before image-to-video. Do not start the Veo, Grok, or Seedance stage until the first frame passes validation.
2. Select exactly one source prompt. Use the natural lifestyle prompt for a warm, non-influencer character whose short description, outfit, and product action come from the user. Use the beauty-presenter prompt for the specified 9:16 Chinese skincare-presenter portrait.
3. Never merge the two source prompts. If the user explicitly requests both, run them as two separate image tasks and preserve each source file independently.
4. For the natural lifestyle prompt, require the user's explicit values for 简要描述人物+穿搭 and 人物与产品的动作. Supply those values as separate runtime inputs; do not edit the stored source prompt or infer missing product use.
5. For the beauty-presenter prompt, require 参考的妆容.jpg, 发箍.png, and 服装.png when the user expects the named reference matching. If any required reference is missing, ask only for the missing file instead of pretending it was supplied.
6. When product scale matters, require a clean real-product reference plus exact dimensions, a size-comparison image, or one user-approved scale frame. Apply the size and refinement supplement; do not infer product dimensions or accept an unapproved generated image as size evidence.
7. Submit the selected prompt to GPT Web with GPT Image 2 together with the user's references and separately supplied facts. Keep product scale, grip, contact points, clothing, hairstyle, face direction, mouth visibility, and lighting consistent with the selected prompt and references.
8. When multiple versions are requested, generate each as a separate 9:16 image. Never return a collage, grid, split screen, or combined image in place of the independent files.
9. Inspect every generated frame at full size. Verify natural facial identity, consistent skin texture, correct fingers and hands, unobstructed eyes and mouth, believable product geometry, correct person-product action, usable framing, and no unintended text or interface elements.
10. Reject and regenerate the still image when any identity, anatomy, product, scale, contact, or lighting error would become visible in motion. A video prompt cannot repair a structurally wrong first frame.
11. After product scale and contact pass validation, keep the approved frame unchanged. When refinement is requested, create a separate same-canvas reference with mosaic only over the product center, then upload the real product reference, approved scale frame, and mosaicked reference as separately labeled inputs. Freeze the product border, apparent dimensions, hands, contact points, person, camera, and scene while refining only the product center.
12. After approval, keep this frame as the character and opening-frame anchor. If face-mask preprocessing is requested, create the separate face-mask reference next; then route to Veo, Grok, or Seedance as requested.

Because this workflow is primarily image-to-video, the first-frame image determines most of the final video's realism. Treat first-frame approval as a hard quality gate, not as an optional preview.

Prepare Face-Mask References

When the user requests face masking or a Seedance workflow needs a processed portrait reference, read and apply references/face-mask-preprocessing.md.

1. Run this module after the character image exists and before video generation.
2. Keep the unmasked image as the identity source and save the mask treatment as a separate reference.
3. Do not upload user media to PromptHub or another external site without explicit authorization.
4. Treat PromptHub as an optional external tool. Do not claim sponsorship, guaranteed availability, or permanent free access.
5. Continue with the immutable Seedance prompt after reference preprocessing is complete.

Route Parent-Child DIY

Apply this route when the user mentions parent-child DIY or equivalent wording in any language. An explicit Seedance request still takes precedence and uses only the Seedance branch.

1. Require a product reference image and a product size or dimension reference. Also require material-kit and picture-instruction references when their appearance cannot be established from the supplied product evidence. Do not infer product dimensions, parts, tools, or instruction content.
2. Use GPT Image 2 for the image stage. Submit the original parent-child DIY prompt together with the user's references and separately supplied product facts.
3. Generate the three A-version scenes as three separate 9:16 image files, one scene per generation. Never request or accept a collage, grid, split screen, triptych, contact sheet, or single combined image as the three-image deliverable.
4. Keep the same mother, daughter, clothing, home craft room, DIY tools, and product identity across all three images. Keep product size and design locked to the references.
5. Never place a product box on the table. The table may show only the referenced material packets, picture instruction manual, tray, and necessary DIY tools. Scene 2 must not contain the completed product.
6. If the user explicitly requests a B version, generate a second set of three independent images. Change only camera position, natural gestures, and tabletop prop placement; keep every locked identity and product fact unchanged.
7. After the images are complete, read and use references/veo-grok-multi-scene-prompt.md verbatim to compile the video prompt. Treat image 1 as the opening frame, image 2 as the middle-scene visual reference, and image 3 as the ending frame.
8. Use mode E because the completed-product state, instruction-reading state without the completed product, and final presentation state require explicit scene changes. Write exact timed hard jump cuts while keeping each scene's internal camera motion continuous.
9. For a single start-and-end-frame generation, supply image 1 as the first frame and image 3 as the last frame, while locking image 2 into the timed middle scene. For separate video units, state the first and last frame of every unit explicitly and use only generated or user-supplied frames.
10. Produce the final prompt for Veo 3.1 or Grok as requested. Do not rewrite either bundled source prompt while connecting the image and video stages.

This route overrides the generic Veo/Grok direct route, the post-script product-display route, and the unboxing route unless the user explicitly requests separate additional deliverables.