slide-design-skill CLAUDE.md

slide-design-skill — Session Rules (READ FIRST, applies to EVERY session in this repo)

The product has NO selectable styles. None. Ever.

There are no themes, no templates, no presets, no "flagship skills" the user picks from. The user describes what they want; a style is generated/derived from that brief, every single time. "Choosing a style" is not a concept that exists here.

NEVER surface internal machinery to the user

The folders under skills/ are internal generator seed/reference material only. The user must not know they exist. When talking to the user, you are FORBIDDEN from saying (or implying) any of:

• "flagship skill(s)", "preset(s)", "template(s)", "theme(s)"
• the name of any folder under skills/ (consulting, pitch, stripe, …)
• "bespoke X (not chosen from the presets/flagship skills)" ← this exact leak just happened, never do it
• "I matched the built-in …", "deriving from stripe.com as the brief says", or any narration of how the style got resolved

If you catch yourself explaining the approach or the architecture, stop. The user does not want a status report on the plumbing.

What to actually do

1. Take the brief (topic + content + look-in-any-form).
2. Silently derive the look and generate a fully bespoke style for this deck (its own tokens, slide types, and chrome.css).
3. Author the deck, render it, show the user the result (screenshots / the HTML).
4. Iterate on the user's feedback.

Minimal narration. No menus. No meta-talk. If the brief carries zero signal about the look, ask ONE short question about the desired look — otherwise never ask, just build.

Moodboard step (image-first style anchor, optional but preferred)

When the look is open (inspiration-style briefs, weak signal) and FAL_KEY is available: generate 2 moodboards via composeMoodboardPrompts(subject) (engine/moodboard.ts) rendered through FalProvider (flux/dev), show them to the user, let them pick. The prompts already rotate each board onto a different unexpected colour axis — NEVER write moodboard prompts by hand without an axis; the image model carries the same genre-default bias as the LLM and will return the cliché palette (verified: "premium chocolate" → beige+espresso every time). View the approved board yourself, extract palette hexes / type mood / material world, and feed it into the generation brief via moodboardDirectionBlock(...). An approved board outranks the banned-default-palette rule (the client chose it). A concrete image to translate produces visibly more coherent styles than adjectives (veta A/B, 2026-06).