Pixel Pal Carousel

Pixel-Pal Carousel

An educational-carousel design system reverse-engineered from a high-performing reference deck. Orange pixel-art mascot + clean editorial typography + hand-drawn accents. Legible, scroll-stopping, and repeatable slide-to-slide.

READ references/design-system.md FIRST — it is the source of truth for the exact hex palette, the two slide archetypes, the mascot spec, the accent library, and the locked prompt template. This SKILL.md is the workflow; that file is the look.

When to use

• "Make a carousel in this design/format" (with the pixel-blob reference)
• Educational / how-to / product-launch / explainer carousels for IG or LinkedIn
• Any deck that should match Samin's channel: clean, light, orange accent, playful mascot

Model & format (LOCKED)

• Model: gpt_image_2 (GPT Image 2.0). It is the ONLY model here — it renders exact text, layout, and typography. Never substitute Soul/Nano/Seedream for text-heavy slides.
• Aspect ratio: gpt_image_2 does NOT support 4:5, so use 3:4 (its closest supported portrait ratio) for IG carousels. (Use 1:1 only if the user explicitly wants square.)
• One higgsfield_generate_image call per slide. For an N-slide deck, fire N calls IN PARALLEL (respect the workspace text2image concurrency cap — default 4; submit ≤cap, poll higgsfield_job_status to terminal, then submit the next batch).

Workflow

1. Get the content. From the user's topic, transcript, URL, or notes. If a YouTube URL is given, pull the transcript/metadata (youtube-content / youtube-research / video_analyze) for concrete facts and talking points.

2. Write the slide plan (before any image). Decide slide count (5-7 is the sweet spot) and for EACH slide draft: archetype (HOOK vs INTERIOR vs CTA), section chip number+label, headline (mark which ONE word/phrase is orange), body copy, mascot pose+props, and any tip-box / speech-bubble text. Keep ONE idea per slide. Slide 1 is ALWAYS the dark serif HOOK. Last slide is usually a CTA/recap.

3. Build each prompt from the LOCKED template in references/design-system.md. Fill the placeholders; do NOT paraphrase the style block — it must be byte-stable across slides so the mascot and palette stay consistent. Put every literal string the slide must show in "double quotes" and add "spelled exactly, verbatim".

4. Generate all slides in parallel with gpt_image_2, aspect_ratio:"4:5". Call higgsfield_generate_models_explore(action='get', model_id='gpt_image_2') once first if you need live params.

5. QC pass (MANDATORY). When jobs finish, run image_analyze on the full set and check: (a) every headline word spelled correctly, (b) the orange word is the RIGHT word, (c) mascot is consistent orange pixel-art across slides, (d) section chip numbers are sequential, (e) no cropped text, (f) tip-box present where planned. Regenerate any slide that fails (retry same prompt once; then simplify).

6. Deliver. higgsfield_upload every final slide and present them in order as MEDIA:, slide 1 → N. Offer the caption + hook line.

Pitfalls

• GPT Image sometimes renders long body paragraphs with typos — keep body copy short (≤ 2 short sentences) and always QC + regenerate.
• If the mascot drifts in color/style between slides, restate the mascot spec block verbatim and add "identical 8-bit pixel-art orange blob mascot as the other slides".
• Don't put more than ~7 words in a headline — it shrinks and loses impact.
• Keep exactly ONE accent color per slide besides orange (the reference uses blue ONLY for the "pictures/words" idea). Orange is the primary accent everywhere.
• The dark HOOK slide uses a SERIF headline; interior slides use BOLD SANS. Do not mix.