Layer 3
Meta-preferences: how you want it to behave
Not what it tells you, but how. Length by context, language by domain, opinion assertiveness, citation style, your programming language. The tactical decisions you'd make in your first conversation with a new human assistant.
What it captures
Six canonical keys: length by context (concise, medium or detailed by question type), language by domain (code in English, dental pain in your mother tongue), contradiction tolerance (how much pushback you allow), opinion assertiveness (how much it commits when stating a view), citation style (inline, footnote, APA, MLA or none), and main programming language.
Why it matters
Every miscalibrated preference is accumulated friction. An AI that gives you five paragraphs when you wanted one, or that caves at the first counter-argument when you wanted a real argument, doesn't fail at the big things — it fails at the everyday things. These six keys cover 80% of that friction with three minutes of setup.
How they're configured
Direct dashboard form, no Proust, no detours: each key has three or four predefined values. Pick, save, get on with your life. If something doesn't fit later, change it in thirty seconds. The AI recompiles the prompt and adapts.
Privacy and plan
Available across all three plans (Essential, Premium, Professional). The block injects before the cognitive portrait, capped at 150 tokens. You can clear any preference at any time.
Frequently asked
- Can preferences depend on the topic?
- Yes. Language by domain, for instance, lets you say "technical in English, personal in your mother tongue". Length too: concise for data, detailed for reasoning.
- What if I don't set any?
- The AI uses sensible defaults: medium length, balanced assertiveness, inline citations when relevant. Preferences are an accelerator, not an obligation.
- Are they respected on every platform?
- On afini.ai, yes. If you export your portable PCP and use it with another external AI, it depends on how well that AI honors the injected block. Most modern models respect it reasonably well.