Brand-style ingestion into Claude Design (2-month maturation)

Designers collect brand assets (logo specs, palette, type stack, examples of accepted and rejected past work) and feed them into Claude Design (or equivalent vision model) over many sessions until the model produces on-brand work. Initial outputs are visibly broken ("6-fingered hands; extra legs"); after ~2 months of iterated training and prompt-refinement, franchise customers can't distinguish AI-produced assets from designer-produced ones. Designers shift their time to high-end work.

The 2026-Q1 ecosystem changed the entry point. claude.ai/design now accepts a Figma design system or React kit as input and one-shot generates landing pages, decks, and design-system artefacts; the Anthropic frontend-design skill bakes design-thinking guidelines into Claude Code; Magic Patterns + Claude Code via MCP offers a richer component palette; and the Google AI Studio → Claude reintegration dance handles aggressive style restyles. The 2-month maturation pattern below still applies on top of these tools — they reduce the floor-of-quality, not the time-to-on-brand. For the canonical 2026 framing, see the new Tier A case on AI-driven design ( ). This sub-case stays focused on the patient training loop that gets a specific brand to ~80% AI-produced.

A small franchise replaced ~80% of routine asset production (social posts, franchisee-facing materials, deck templates) with this workflow at $200/month total tooling cost — versus the prior path of either freelance design queues or in-house designers spending hours on routine work. Plan to productize: replace an external social-media service with a lower-cost bot service for franchisees.

1. Brand-style packet: hi-res logo, palette swatches, type spec, 30+ approved examples, 10+ rejected examples (with annotations).
2. Install the Anthropic frontend-design skill in Claude Code; it materially improves the floor-of-quality of generated landings.
3. Designer onboarding session — load packet, walk through approved/rejected pairs.
4. Designer iterates: produce → designer rejects → designer adds rule to brand-style packet → repeat.
5. Track 3 quality bars: structural correctness (no extra fingers / mangled text), brand correctness (palette / type), aesthetic correctness (composition).
6. Burn through ~3 weekly quotas of the $200 plan during the ramp; budget the cost as part of the project.
7. After ~2 months: designers spot-check rather than produce. Routine work goes through the model.
8. Maintain the brand-style packet as a living file in a private repo.
9. For one-shot artefacts (decks, single-page landings), feed the brand-style packet into claude.ai/design along with your Figma design system or React kit; iterate via prompts; export.
10. For aggressive restyles, use the three-screen dance (Claude Code → Google AI Studio → Claude Code) — fold all components into one file, restyle in AI Studio with a one-line brief, reintegrate.

• Don't ship the model's output to customers without designer spot-check during the first 2 months. The "6-fingered hands" stage is real and will reach a customer if you skip the gate. After maturation the gate becomes a 30-second spot-check.
• NEW — Naked Claude produces generic landings. Without the frontend-design skill, output looks the same as every other AI landing. Install the skill before complaining about quality.
• NEW — AI-Studio restyle can produce "crypto-aesthetic flashy". The three-screen dance favors maximum surface polish + animations. Good for some brands, wrong for editorial / minimal taste. Prompt for the aesthetic explicitly.