Virtual Influencer Slideshows

Virtual Influencer Slideshows

Produce post-ready TikTok slideshow decks for an existing virtual persona, end to end. The character's identity kit must already exist (built by the create-virtual-influencer skill).

What you need from the user

• Which character to produce for — the folder name — and how many decks to produce.
• The product/app to promote and its EXACT positioning — what it does, what it must NEVER be described as, and whether it is paid, free, or platform-limited — or "none" for no product placement.
• The base working directory. Follow the conventions established there.

First read the character's persona doc, voice pack (with its banned-tells lint), and world bible; do not work from memory.

Expertise and tools

Adopt the expertise of a short-form content strategist and an AI-image art director. Use the tiktok-slideshows and persona-image-curation skills. Tools: an Apify Pinterest scraper (to extend the scene library), an image-generation tool that accepts reference images (for character slides only), a deterministic HTML/CSS-to-image text renderer run through headless Chrome (for all slide text), and ImageMagick.

Steps, in order

1. Set up the render pipeline in the character's slideshows/ folder: copy the deterministic text renderer, its fonts, vendor libraries, and logo chips; create a fragments/ folder; confirm the renderer auto-merges per-deck fragment files; render one throwaway test slide to prove the pipeline works before producing anything.

2. Plan formats and topics. Read the format-interaction research and choose a mix leaning on the highest-interaction formats (product roundup, tutorial, listicle), reserving story/POV for relatability. Define the requested number of topics in THIS character's lane, each distinct and none overlapping the topics of any other persona in the project. Assign each a format, and mark the small subset that will carry a product mention.

3. Extend the scene library if needed for archetypes or depth the pool lacks: scrape candid-biased first-person queries, triage by dimension forensics, download, run parallel full-resolution persona-test reviews, and merge keepers into the pool with manifest updates. Use only verified-real phone photos.

4. Partition the pool into an exclusive background allocation per deck — no background is ever reused across decks. Record the allocation to a file.

5. Produce each deck (you may fan out a team of subagents, each owning a few decks and modeling one finished reference deck, and each banking every deck to disk before starting the next). Per deck:

   • (a) Script in the character's voice — payoff slide first, then two alternate hooks, one idea per slide, strongest point on slide two, a save CTA. Every tip must end in an exact script or a right-versus-wrong contrast. Then lint every line against the voice pack.
   • (b) Backgrounds: copy the assigned backgrounds into the deck folder in slide order and re-run the full-resolution persona test on each; drop and note any that fail rather than substituting from outside the allocation.
   • (c) Character slides: generate only the slides the character appears in (usually the hook and the payoff/intro), text-free, attaching the face sheet as the identity anchor plus the chosen outfit reference and any prop sheet whose object appears. Describe only scene, pose, camera, and which outfit/hair/earring/makeup option — never re-describe the character's identity. Copy each generated image into the deck folder immediately after generating it. Verify at full resolution: two arms, two hands, five fingers each, signature accessories present, identity and canon props correct; regenerate on any doubt. Rotate the outfit/look across decks.
   • (d) Text: write the deck's copy into fragments/<slug>.json and render it deterministically — never bake slide text with the image model.
   • (e) Review: montage the rendered slides and inspect at full size; fix wraps, subject collisions, contrast; re-render.
   • (f) Red-team every slide as a hostile commenter hunting authenticity tells: hands, in-scene text, prop continuity against the world bible and every other deck, shadow physics, calendar truth for the posting date.
   • (g) Document: write the deck's index.md with both hooks, slide-by-slide copy, design notes, an assets list giving each image's source and persona-test verdict, a caption with hashtags, audio direction, and one continuity-ledger row.

6. Final QA across all decks: montage a full grid and inspect identity consistency; check cross-deck continuity of outfits, props, and the pet; merge every deck's ledger row into the world bible; write the slideshows index; update the project index chain; and send a completion notification.

Done means

The requested number of decks each exist as numbered <NN>-<slug>-final.jpg files, each with a fragment and a documented index; every image passed the persona test and red-team pass; every line passed the voice lint; any product mention is accurate, singular per deck, and capped with one true limitation; no background is reused; and nothing contradicts the world bible or any previously produced deck.

Hard rules

• Generate scene images only for slides the character appears in; every other slide is a verified-real pool background used as-is.
• Render all slide text deterministically through the HTML/CSS renderer, never through the image model, and never use the renderer's vignette.
• Keep the character's phone, jewelry, nails, hair, pet, and other canon objects exactly as the world bible specifies; carry identity from the reference sheets, never from prompt text.
• Write only in this character's voice; never borrow a register the voice pack marks as belonging to another persona.
• If a product is placed: describe it only by its true positioning, never by what it is not, never call it free unless it is, and never give it a flawless review — cap it with one honest limitation (at most 8/10).
• Plan for two operational hazards: the image-generation service can exhaust its credits mid-run — bank each completed deck to disk before starting the next, and fall back to a text-forward hook over a real background when a character slide cannot be generated; and the image tool's output directory may be volatile — copy every generation into the deck folder immediately.
• If a step fails twice, note it in the deck index and move on rather than stalling.