Global Instructions & Behavioural Rules

What You Need to Know

Domain 5 has covered project-level configuration (5.1), persistent context files (5.2), and memory (5.3). This lesson zooms out to the account-wide layer: the settings that apply to every conversation, every project, and every chat — the rules that are always on.

Understanding where to put each rule — Global vs. Project vs. Style — is one of the most heavily tested concepts in Domain 5. Put a rule in the wrong layer and it either doesn't apply when needed or applies when it shouldn't.

Profile Preferences: the company handbook

Profile Preferences live in Settings > Profile and apply to every conversation across the entire account. Think of them as the company employee handbook — it doesn't matter whether you're in the Marketing department or the Finance department; everyone follows the same core rules.

This is where you put things that should never change regardless of context:

• Your name and how you want to be addressed
• Your language preference (e.g., "Always use British English")
• Universal formatting rules (e.g., "No preambles or flattery")
• Behavioural guardrails that apply everywhere (e.g., "If I ask about a product or statistic, always use web search first")
• A timestamp ("Instructions last updated: April 2026")

Behavioural guardrails: hard rules, not suggestions

The most valuable Global Preferences are explicit "stop doing this" rules that constrain Claude's default behaviour. Claude's factory settings include:

• Agreeable preambles ("Great question!" "That is an excellent point!")
• Hedging language ("It might be worth considering..." "Perhaps you could...")
• Validating everything the user says
• Verbose responses that pad content with filler

Guardrails override these defaults:

• "No preambles, no flattery — start every response with substance"
• "If my logic has a gap, say so directly rather than agreeing"
• "Always web-search before giving advice on pricing, statistics, or current events"
• "Maximum 3 sentences per paragraph"

The distinction between soft and hard rules matters enormously. "Try to be concise" is a soft command that Claude treats as a suggestion — it will produce 500-word responses when it deems it appropriate. "Maximum 50 words per response" is a hard rule that produces 50-word responses. Use concrete numbers, absolute constraints, and specific rules rather than aspirational guidance.

Custom Styles: formatting without substance

Custom Styles — "Concise," "Explanatory," "Formal" — are formatting presets that control how Claude delivers its response. They affect sentence length, use of headers, level of detail, and overall tone of the output. You can toggle between Styles without changing the underlying instructions or knowledge.

The critical distinction: Styles and Instructions are different things. Instructions define what Claude knows and how it behaves. Styles control how the final output is formatted and presented. A Style can't enforce a Risk Rating header at the top of every response. A Style can't require web search before answering. Those are behavioural rules that belong in Instructions.

Think of it this way: Instructions are the recipe. Styles are the plating. The same dish (analysis) can be served on a minimal plate (Concise) or an elaborate one (Explanatory), but the ingredients and cooking method (Instructions) don't change.

The layered context stack

Claude processes context in layers, and understanding this hierarchy is essential for troubleshooting:

1. Global Preferences — load first as the foundation for every conversation
2. Project Instructions — add task-specific context on top of the global layer
3. Custom Style — adjust the final delivery format of the response

When layers conflict, the more specific layer generally takes precedence. A Project Instruction that says "Use casual language" will override a Global Preference for "Formal tone" within that project. This is by design — project-level specificity should override account-level defaults when they are in direct conflict.

If a response ignores a rule you expected to apply, troubleshoot by checking which layer the rule lives in:

• Is it in Global but the project has a conflicting instruction?
• Is it in one project but you are chatting in a different project?
• Did you add it after starting this conversation (remember: Global Preferences only apply to new chats)?

Temporal context: instruction versioning

Adding a timestamp to your instructions — "Instructions last updated: April 2026" — serves two purposes. First, it gives Claude temporal context about when your rules were written, which helps it avoid applying outdated guidance. Second, it reminds you when instructions need refreshing. As your role, company, and tools evolve, your instructions should evolve with them.

---

Common Mistakes

---

Hands-On Activity

---

Practice Question

---

Sources

• Understanding Claude's Personalization Features — Anthropic
• Claude Cowork Guide 2026 — FindSkill.ai
• Get started with Cowork — Anthropic