AI Personal Preferences

Better output usually comes from better behavior design, not a cleverer prompt. Personal Preferences is where you do that design — once — instead of re-introducing yourself every single conversation.

#What it is

Personal Preferences is a free-text field at Settings → General → What personal preferences should Claude consider in responses? on Claude.ai. Whatever you write there is applied automatically to every new conversation — no re-explaining.

Most people stop at typing prompts. But each new chat starts blank: Claude doesn't remember who you are, what style you like, or what field you work in — so you retype the same "about me" every day. Preferences removes that tax.

#How it works

When you open a new conversation, your saved preferences are injected ahead of your message:

[Your saved Personal Preferences]
        -> injected into
[Claude's system prompt]
        +
[Your conversation]
        ->
[Claude responds with full context]

• Automatic — nothing to do per chat; it's always there.
• High priority — it sits at the front of the context, read before anything else.
• Global — applies across web and mobile.
• Costs tokens — every word is consumed in every conversation, so keep it tight.

#Preferences vs. Memory

They look similar but do opposite things:

#Setting it up

On claude.ai: open your account menu → Settings → General, find the preferences field, write it, and save. Every new conversation from then on uses it.

#Six kinds of preference

Useful preferences fall into six groups. The last one is the highest-leverage:

• Identity — who you are, so answers are framed for you (role, stack, what you're building).
• Behavior — kill the defaults you dislike.
• Format — how output is presented (brevity, bullets, code comments).
• Domain — your expertise level, so depth is calibrated.
• Language & style — language and tone.
• Process — how Claude should approach the work. This changes behavior the most.

Behavior and process rules are where the real gains are:

# Behavior
- Don't open with "Certainly!" or "Great question!"
- Don't summarize what I just said before answering
- Challenge my reasoning if you see a hole
 
# Process (highest impact)
- Search the web before recommending a specific library/version
- If a task splits into >=2 independent parts, run them in parallel
- Always finish with a synthesis step
- Priority: Correctness > Clarity > Speed

#Best practices

#Where it sits in the stack

Claude has four configuration layers; higher = more general, lower = more specific and can override:

#Limitations

• Token cost — embedded in every conversation, eating context from the start.
• Not portable — Claude.ai preferences don't auto-apply to direct API use.
• Not a hard command — Claude can deviate when the conversation demands it.
• Not a substitute for context — for codebase rules use CLAUDE.md; for big reference material use Project Knowledge; for complex reusable behavior use Skills.