Brand DNA. What each field does
Brand DNA is the configuration every Siren generation respects. It is not a prompt prefix you hope the LLM honors. It is a set of structured fields, enforced server-side. Slogan and disclaimers are locked at the renderer. The LLM cannot author a different line.
This page covers each field, what it does, and where it surfaces in generation.
Identity
Brand name
The name as you want it to appear on cards, videos, and signatures. Used in wordmark slots across templates. Editable. The renderer caches it per generation.
Logo (light and dark variants)
PNG, JPG, or WebP, up to 2 MB. The dark variant is used on light backgrounds. The light variant is used on dark surfaces. If only one is provided, Siren falls back to a text wordmark on the surface where the missing variant would have been used.
Note. Use a high-resolution PNG at 2× your intended display size. SVG upload is not available in v1. It is planned for v1.5 with server-side sanitization.
Slogan
For Siren itself, the slogan is server-locked: Siren · a thousand voices, one signature. For your brand, set your own one-line slogan. The renderer enforces whatever you set. The LLM cannot author a different one.
Product screens
Named screenshots of your product (home, feature, dashboard, and so on). Phone, desktop, or tablet. Light or dark UI. Stored on R2 and reused on device-framed masters.
An agent can upload them over MCP (upload_product_screen) or the key API (POST /api/v1/programmatic/brand-assets/screens) with an image_url or image_base64. Same for get/patch of overview, tagline, offer, and features (get_brand_dna / update_brand_dna). Brand Memory is not on that surface.
Voice
Tonal archetype (9 choices)
Pick one of nine: Trustworthy, Builder, Aspirational, Playful, Authoritative, Contrarian, Empathetic, Pragmatic, Edgy. Each is a bundle of tone-of-voice rules, motion presets, and copy heuristics that the Reasoner respects.
You can change this at any time. Existing drafts stay as rendered. Future generations adopt the new archetype. Brand Memory entries from the prior archetype still inform the new one. Voice is layered, not switched.
Banned words
A list of words and phrases the LLM must never produce. Enforced at two stages. The Reasoner refuses to draft them. The renderer post-processes generated copy to flag any that slip through.
- One per line
- Case-insensitive
- Substring-matching by default. "synergy" blocks "synergistic" and "synergies"
- Wrap in slashes for exact-word matching:
/synergy/matches only the standalone word
Required disclaimers
Text that must appear on every regulated asset. Server-locked at the renderer. The LLM does not author this. It is stamped onto the rendered output.
Common patterns:
- Financial: "Not financial advice. Past performance does not guarantee future results."
- Crypto: "This is not an offer to sell securities. Risk capital only."
- Health: "Consult a qualified provider before making decisions about your health."
Why server-locked. Disclaimers are the most common drift point in LLM-generated content. Siren does not trust the LLM to include them. They are rendered at the template layer, where the LLM has no write access.
Visual
Colors
Three slots: primary, accent, optional secondary. Primary is the dominant color. Accent is for spotlights and CTAs. Secondary is occasional contrast.
Colors are the most-used Brand DNA field. Every template references them. The Tweaks panel can override accent per generation without re-running the LLM (Phase 1.5 added accent_color_override; saturation and surface finish are queued for v1.5).
Fonts (display and body)
Two slots: display font for headlines, body font for everything else. Pick from the curated list (a Google Fonts subset) or upload a custom .woff2 on Pro+.
Custom font upload requires a license file the legal team can verify. Siren does not ship unlicensed typography.
Audience
Audience parameters
Optional. Leave the fields blank and Siren writes for a general audience until your first 30 dispatched posts return engagement data. From then on Hits (real views, likes and replies from posts that shipped) shape what runs next.
How Brand DNA flows through generation
- Prompt arrives. Siren reads your Brand DNA and the prompt together.
- Video. Siren picks one of its video types that suits the prompt and your archetype, writes the words into it, and refuses banned words. The video renders from code with your colors, fonts, and logo. Disclaimers are baked in by the renderer, not the LLM.
- Still. Siren writes the picture from your prompt, Brand DNA and Brand Memory, and an image model paints it in your palette, with your product screenshots when the prompt calls for them.
- Caption. Written in your voice, checked against banned words.
Every step has a Brand DNA gate. Drift is structurally hard, not "trust the LLM" hard.
Updating Brand DNA
You can update fields at any time. Changes take effect on the next generation. Existing assets stay as rendered, so you can compare. Brand Memory entries persist across DNA changes. Voice accumulates through edits, not resets.