Most "brand settings" in content tools are a text box. You paste a style-guide paragraph. The model reads it once per session and then loses it under pressure from the rest of the prompt.

Brand DNA in Siren is a structured schema. It is enforced at more than one layer, and it is applied before copy is written, not after.

What Brand DNA contains

Brand DNA is not a free-form style guide. It is a typed profile. Different fields feed different parts of the pipeline.

Archetype. You pick a tonal archetype (Trustworthy, Builder, Aspirational, and the rest of the fixed set). That choice sets the register the Reasoner aims for, and it shapes how the Reasoner checks impact. A Builder brand is expected to sound pragmatic and forward. A Trustworthy brand is expected to sound measured. Drift off-archetype is checked as a first-class signal, not a soft preference.
Palette and visual identity. Brand hex codes, accent colors, font personality. The render engine uses these to constrain color fills and type. The palette does not change the Reasoner. It does change every rendered output. Visual drift erodes recognition the same way voice drift does.
Tagline. A locked phrase, or a short set of phrases, that anchors the voice. The Reasoner uses the tagline as a tonal reference, not as a quote to insert. If the tagline is punchy and direct, generation is aimed at that register. The tagline is never appended automatically.
Banned words and topics. An explicit exclusion list. Banned words do not appear in generated copy. Banned topics are subjects the Reasoner must not frame content around, even obliquely. This is a hard constraint. If a prompt tries to generate content that touches a banned topic, the Reasoner rejects the run before render completes.
Tone axes. Dial values (formal versus conversational, serious versus playful, direct versus exploratory). The axes are continuous, not binary, and they interact with the archetype. The same playful setting on two different archetypes does not produce the same register.

How each field is applied

The Reasoner is a two-stage process. The first stage produces a SirenSeed: a structured JSON object that defines the generation intent (angle, hook, supporting points, tone target, archetype alignment, predicted angle theme). The second stage renders that seed into copy.

Brand DNA is applied at the SirenSeed stage, not at the copy-rendering stage. If you wait until copy exists to apply brand constraints, the creative direction is already set. You get text that avoids banned words and still feels off, because the angle and hook were chosen without the constraints.

Applied at SirenSeed, the archetype shapes the angle. The tone axes shape the hook. Banned topics exclude whole directions before a word of copy is written. The copy that comes out is on-brand in structure, not only in vocabulary.

Why a pasted prompt is not enough

The usual approach is: paste brand guidelines into the system prompt. That works for simple cases. It fails in predictable ways.

Context-window pressure. The longer the prompt, the less attention the model gives the guidelines. Brand DNA fields are not prose in a system prompt. They are a structured schema injected into the SirenSeed function-calling interface. The model has to output a typed object, not free text.
Ambiguity at the edges. "Write in a conversational but professional tone" is an average-case instruction. When the prompt asks for a contrarian take, a reply to a competitor, or a launch post with a tight character count, that sentence does not give the model enough to hit. A typed archetype plus tone axes gives it a target it can aim at under constraint.
No external check. A system prompt cannot tell you whether the output actually sounds like your brand. The model grades its own work. Hits after you post are the outside check: real engagement on work that already cleared Brand DNA.
Nothing compounds. Every session starts fresh. Brand Memory stores winning outputs, confirmed by real engagement, as reference. Later generations in your workspace are influenced by what has already worked for that brand and audience, not only by the current prompt.

Why this matters more as you scale

For a single-person brand, the founder's ear is the consistency check. They read every output and reject the ones that feel off. That works at 10 posts a month.

At 100 posts a month, across platforms and people, that ear does not scale. The person who writes Monday's LinkedIn post and the contractor who schedules Friday's X thread are not reading each other's work. Drift accumulates. A year later, the LinkedIn voice and the X voice are measurably different, and neither quite matches the archetype the founder had in mind.

Brand DNA enforced at generation time means that drift cannot start at the generation layer. The archetype is the same for every run, every teammate, every platform. Variation happens inside the archetype (different angles, hooks, formats), not across it.

See how Brand DNA applies to a prompt at mysiren.ai/how-it-works, or set up your Brand DNA.