Carousel Generator

Carousel Generator Skill — Nope Fallacy

This is the full studio: idea in, finished slide copy out, then a stack of paste-ready image prompts out. The whole point is that Adrita never has to shuffle text between Google Docs and ChatGPT — she copies one prompt block, pastes it into ChatGPT (Image 2 model on), and gets a slide. Then the next block. Then the next.

Carousels are the most underrated format on Instagram: higher engagement than reels, multiple distribution channels (home feed, explore, reels tab when audio is added, and a "second-chance" re-show of slide 2 to anyone who skips slide 1). Every carousel that ships should be engineered, not just drafted.

Brand facts (use these, don't invent):

• Instagram: @adrita.space · YouTube: @nope.fallacy · Branded hashtag: #nopefallacy
• Brand voice = Nope Fallacy by Adrita Mahata. Best-practice voice for now; if references/house-style.md has filled-in winners, match those patterns instead.

The pipeline (4 phases)

Think of this as four phases. Phases 1–2 are copy. Phase 3 is the anchor decision. Phase 4 is prompts. Don't sprint to prompts — bad copy + great visuals = a pretty post that dies. Roughly 40% of the effort is copy, and the rest follows from a strong slide-1 anchor.

PHASE 1  Scope + write the deck      →  slide list, hook variants, per-slide copy, caption, hashtags
   ↓     (Adrita approves / iterates)
PHASE 2  Lock the copy               →  final, numbered slide text she's happy with
   ↓
PHASE 3  Visual reference + direction →  ask for her reference image; write per-slide visual direction
   ↓
PHASE 4  Generate image prompts      →  paste-ready prompts: slide-1 anchor (×5), then each slide (×3)

Stop and get explicit approval at the end of Phase 1 before moving on. The copy is load-bearing — everything downstream inherits its quality.

PHASE 1 — Scope and write the deck

Read references/carousel-bible.md and references/slide-count.md first. They carry the 10 commandments, the playbook skeletons, the slide-3 danger zone, hook formulas, and the slide-count matrix. They're load-bearing, not background.

Then:

1. Pick the playbook from the Bible's 10 (Tip/Value · Story/Transformation · Curation/Listicle · Same-Meme · Quote Collection · Video Postcard · EDC/Product · Two-Slide Hook · Brand Story · Gamified). State which and why.
2. Pick the slide count. Default to the Magic 8 (hook + 6 content + CTA). Use references/slide-count.md for topic-type overrides. IG hard cap is 10.
3. Decide TOFU or BOFU. TOFU = broad, growth, "follow for more." BOFU = narrow, problem-solving, "comment KEYWORD." This changes the hook angle and the CTA.
4. Pick the hook category (Challenge · Mistake · Listicle · Story · Curiosity Gap · Authority) and apply How-I > How-To whenever credibility allows. "How I" is more relatable than "How to."
5. Write the cover (S1) with 3 hook variants. 3–7 words each. Run the hook checklist (in the Bible) on each. Recommend one, show all three so she can swap.
6. Write S2 as a second-chance hook. IG re-shows S2 to people who skipped S1, so it has to stand alone AND deepen the curiosity. Don't restate the cover — reveal something.
7. Front-load the densest value at S3. This is the algorithm checkpoint. If a viewer reads S3 and bails, they still got value and the algorithm still got the dwell signal.
8. Write the middle slides — one breath, one idea each, max 2–3 short lines. Each must pass the screenshot test: could someone share just this slide to their story and have it make sense? If no, split or rewrite.
9. Write the penultimate slide (the payoff) — the "oh shit" / aha that drives shares.
10. Write the last slide (CTA + reward) — ONE specific action matched to the goal. "Comment [KEYWORD] and I'll DM you [deliverable]" beats "follow for more." Use @adrita.space in growth CTAs.
11. Write the caption — pick ONE strategy (Inside Joke · Expansion · Question · Story Hook). Don't repeat the slides.
12. Add 5–10 hashtags, varied sizes (3–4 niche, 2–3 audience, 0–1 branded #nopefallacy). Avoid 1M+ use tags.
13. Run the anti-AI filter from references/house-style.md. Cut anything that reads generic or templated.

Then present using the Phase 1 output format below and ask: "Want to lock this copy in, or should we adjust hooks / count / angle first?"

Phase 1 output format

CAROUSEL — [Topic]

PLAYBOOK: [which] · TIER: [TOFU/BOFU] · SLIDES: X (why) · HOOK: [category] using [How-I/How-To]

S1 (COVER)
  HOOK A: [3-7 words]
  HOOK B: [3-7 words]
  HOOK C: [3-7 words]
  → RECOMMEND: [A/B/C] because [reason]

S2 (SECOND-CHANCE HOOK): [stands alone, deepens curiosity]
S3 (FIRST VALUE DROP): [densest insight, front-loaded]
S4…S(N-1): [one idea each]  ·  note save/share/completion intent per slide
S(N-1) (PAYOFF): [the aha]
S(N) (CTA): [one specific action]  +  sign-off/reward line

CAPTION ([strategy]): [text]
HASHTAGS (5-10): [list]

🎵 Audio reminder: attach music post-export — without it, only existing followers see this.

PHASE 2 — Lock the copy

When Adrita approves, restate the final numbered slide text in a clean block (S1…SN, exact strings). This is the single source of truth the prompts will quote verbatim. No commentary in this block — just the slide numbers and the exact text that must appear on each slide. This is what makes Phase 4 "copy, paste, done" instead of "copy, fix, re-paste."

PHASE 3 — Build the visual brief

Read references/image-prompts.md now — it holds the visual-anchor workflow, the failure mode to avoid, and the prompt templates. The single most important thing in this phase: AI image models default to a flat typographic poster (big serif, paper texture) for "carousel slide," which is why most AI carousels look identical. The prompts must override that default and transfer the reference's real medium. How well that works depends on whether you (Claude) can see the reference — so this phase has two modes.

1. Offer both modes at copy approval. Tell Adrita: "Paste a visual reference here and I'll tailor every prompt to it exactly — or just say you'll attach one in ChatGPT yourself and I'll write universal prompts that pull the look from your image."

   • Mode A (she pastes the reference to you): best quality. You can see it, so extract a concrete Reference Style DNA (medium, palette, texture, mood, composition, type) and bake it into every prompt. If she only describes it in words, build a light DNA from that.
   • Mode B (she'll only attach it in ChatGPT, or gives nothing): the default. You can't see it, so do NOT invent a DNA you'd guess wrong. Use the universal templates that tell ChatGPT to study and reproduce her attached image's style, with the anti-default guardrail. The look comes from her image. No reference at all yet? Suggest 2–3 concrete Pinterest searches that name a medium for this topic.

2. Write the per-slide brief. Mode A: extract the DNA, then invent an on-topic scene per slide rendered in that medium (not a "layout") — the creative leap that stops slides looking generic; if a slide names a tool, render its logo as an object inside the scene, recognizable. Mode B: note just the exact text + a light, medium-agnostic subject idea per slide, phrased so the attached image governs the rendering.

3. Keep text placement loose. Give placement as a hint, never a spec — the old prompts followed exact placement too literally and slides came out stiff. Phrase it like "let the headline sit naturally, large and dominant the way the reference treats its type; don't force an exact position."

Present the brief and confirm before generating the prompts.

PHASE 4 — Generate the image prompts

Now produce the paste-ready prompts using the templates in references/image-prompts.md. The rules that make this actually save time:

• Use the Mode A or Mode B templates from image-prompts.md to match Phase 3. Either way, every prompt bans the default explicitly — do NOT reduce this to a plain typographic poster or paper background; stay in the reference's medium — because that ban is what makes the reference transfer instead of the model falling back to the generic look. Mode A carries the concrete DNA; Mode B tells ChatGPT to study and reproduce the attached image.
• Transfer the style, never avoid it. Don't include "do not copy the composition" — that pushed the model toward the default in the old version. Reuse the reference's medium and composition logic; change only the subject and text.
• Fully populate every prompt. Quote the exact slide text from Phase 2 and drop in the Phase 3 brief. Leave NOTHING for Adrita to fill except attaching the reference image (slide 1) or the anchor (slides 2+). If she has to edit the prompt, the skill failed its job.
• Slide 1 first, as the visual anchor. Output ONE prompt for 5 versions of slide 1 with her mood reference attached. Slide 1 locks the look — she picks the winner (or regenerates) before anything else.
• Then every other slide, ×3 versions each, each prompt instructing the model to use the chosen slide 1 as the visual anchor (Mode A also reinforces the DNA), so the deck stays one family.
• Show the DNA block once at the top (Mode A only), then format every prompt as its own copy-block (### SLIDE 1 — anchor prompt, ### SLIDE 2 prompt, …) so Adrita can copy one block at a time. Above the stack, give the one-line order of operations: generate slide 1 → pick the winner → attach it as the anchor for the rest.

After the prompts, add the close-out reminders: attach audio post-export; the 95% fix (if one slide is almost perfect, rebuild it as editable layers in Canva Magic Layers / Magic Studio rather than regenerating); and the pre-publish checklist pointer in the Bible.

When to push back instead of drafting

• News/announcement with no angle → ask for the contrarian take or practical use. News alone = skip.
• Too thin for 6+ slides → suggest a single image or a Two-Slide Hook instead.
• Too big for 10 slides → split into two linked carousels, or escalate to a YouTube video (@nope.fallacy).
• Pure opinion with no demo/data/example → ask for one piece of evidence to anchor it.

Reference map

• references/carousel-bible.md — the operating layer: 10 commandments, universal agreements, 10 playbook skeletons, hook formulas, content-slide templates, CTA-by-goal, captions, hashtags, advanced effects, checklists. Read every draft.
• references/slide-count.md — slide-count picker by playbook and by raw topic type.
• references/image-prompts.md — visual-anchor workflow + the slide-1 and per-slide prompt templates. Read at Phase 3.
• references/house-style.md — voice rules (dial up / cut on sight) + anti-AI filter + a fill-in-later "your winners" section to model Adrita's own top performers once she has them.