node-carousel
Free Instagram carousel generator for Claude Code. Procedurally varied, quality-locked, zero paid APIs.
Turn a topic into a branded 5–8 slide SVG carousel in one command. Runs entirely on your Claude Code plan — no OpenRouter, no Gemini, no Canva, no subscriptions.
What makes v0.4 different
Most AI carousel tools are one of two things: dumb text templates that all look the same, or "AI generates the whole SVG" products that look gorgeous in the demo and break in week two.
This one is neither. v0.4 is a procedural design system — not a template engine, not a freeform generator.
- Variety within quality. Every carousel is seeded by
hash(brand.handle + topic + version). The same inputs always produce the byte-identical deck. Different inputs produce visibly different compositions — different typography scale, different decoration picks, different background variations — within a locked aesthetic. Two users generating two topics never get the same deck. Re-running on the same topic always does.
- Token-driven polish. Typography, spacing, grids, color roles, and decoration rhythm come from shared token files (not hardcoded per-slide). Change one token, every slide updates. Swap one preset, the whole deck re-skins without touching the content.
- 8 compositional patterns, 6 aesthetic presets. Cover, list (bullet / numbered), stat-dominant, quote-pulled, split-comparison, CTA — each pattern has multiple typographic + background variations. Presets pick from well-known brand languages (Stripe / Linear, NYT, Pentagram, Vercel, Fontshare).
- Deterministic seeded sampling across 6 variation axes. Size (compact / standard / oversized), weight contrast, kicker presence, decoration flavor, background variant, noise texture — each is a controlled axis. The seeded RNG picks a coordinate in that space. Output feels hand-designed but is reproducible and version-controllable.
What's new in v0.8
Scan as design extractor — not just preset matcher. Prior versions mapped your scanned site to one of 6 canonical presets (editorial-serif, neo-grotesk, etc.) with the brand's accent color overlaid. v0.8 extracts the actual design system — pixel-sampled colors, glow/effect parameters, composition measurements — and renders those directly.
- Pixel sampling (
scripts/sample-pixels.mjs): 13-point grid samples the hero screenshot, decodes PNG via stdlib zlib, clusters dominant colors with ΔE. No new deps.
- Glow detection: identifies bright saturated accent regions from pixel data, records color + position.
- Vision fingerprint (
prompts/vision-fingerprint.md): structured measurements of background composition, focal elements, effect parameters, atmosphere — replaces the old 6-enum classification for scan-first synthesis.
- New
visual.background.type: "scanned": data-driven background recipe. Renderer composes SVG filters dynamically from 4 overlay types (starfield, vortex, blob, grain) + glow. Documented in docs/brand-profile-schema.md.
- Synthesizer Phase 0.75: scan-first synthesis when pixel sampling + vision fingerprint confidence >= 0.7 AND no forced preset. Preset becomes fallback for low-confidence scans.
Effect on real brands
Scanning theproducerschool.com now produces a cosmic-dark recipe with flanking blue glow + subtle starfield — matching TPS's actual hero aesthetic. Prior versions would have picked neo-grotesk preset + TPS blue accent on flat dark bg, losing the cosmic signature.
Forcing preset-first
--preset <name> still short-circuits scan-first. Paper-aesthetic editorial-serif is still reachable: /node-carousel:scan https://yourbrand.com --preset editorial-serif.
What's new in v0.7.1
