Claude Instructions, Styles, and Skills: A Practical Setup Guide

What Changed: Styles, Preferences, and Skills in Late 2026

If you set up Claude a year or more ago, the names have shifted. "Profile preferences" lived under account settings, and a separate styles menu let you pick presets such as Concise or Formal or build custom styles from a description or writing sample.

Anthropic's current personalization article describes three features instead: Instructions for Claude at the account level (found under Settings > Your Claude, or Account in the older layout), project instructions, and skills. Anthropic's help center has also described custom styles moving to skills, with migrated styles appearing under Customize > Skills. If you still see a styles menu in your account, the rollout may not have reached you yet. Check the help center for the current state rather than relying on older tutorials, including this one.

The practical upshot is good: instead of a style being a flat block of formatting rules, it can now be a skill that Claude loads only when it is relevant, alongside other capabilities.

Which Layer Should Hold What

For a deeper comparison of these building blocks, see Claude skills vs projects vs MCP vs custom instructions.

Writing Account Instructions That Help

Account-wide instructions apply to everything, including quick questions where heavy rules get in the way. The best ones are short and true in every context. Anthropic suggests covering things like your preferred methods, common terms, typical scenarios, and communication preferences.

A compact example that stays useful across very different chats:

I'm a product manager at a B2B SaaS company. I work in US English
and metric units.

- Lead with the answer, then the reasoning.
- If my request is ambiguous, ask one clarifying question before
  writing anything long.
- Use tables for comparisons of three or more options.
- Don't add disclaimers unless there is a real risk I'd miss.
- When you're unsure of a fact, say so plainly instead of guessing.

Before and after for a common mistake:

• Before: 900 words including the company's full brand voice guide, three customer personas, and a list of banned words for marketing copy.
• After: the eight lines above at the account level, with the brand guide and personas moved into a marketing project or a writing-voice skill.

The "before" version makes a quick SQL question come back sounding like a landing page. The "after" version keeps everyday answers clean and loads the brand voice only where it belongs.

Setting Up Project Instructions

Projects group related chats with shared instructions and knowledge. Anthropic notes that chats within a project do not share context with each other unless that information is in the project knowledge base, so the project setup is what carries continuity.

1. Create one project per ongoing body of work, not per topic you are curious about.
2. Write instructions that say who the work is for, what good output looks like, and any workflow rules.
3. Add reference files to project knowledge rather than pasting them into instructions.
4. Keep instructions short. Anthropic's tips for conserving context include keeping project instructions concise and removing unused project files.
5. Revisit after a few weeks. Delete rules Claude follows anyway; sharpen rules it keeps missing.

For a full walkthrough, see how to set up a Claude project knowledge base.

Replacing Styles With Skills

Skills are folders of instructions, and optionally scripts and resources, that Claude loads when a task calls for them. Anthropic lists them as available across Free, paid, Team, and Enterprise plans, with code execution required. You manage them under Customize > Skills, and simple ones need no code: the instructions are written in Markdown.

A writing-voice skill that does what a custom style used to do might contain:

---
name: plain-product-voice
description: Use when writing customer-facing product copy,
  release notes, or help articles.
---

# Plain product voice

- Short sentences. One idea per sentence.
- Second person ("you"), active voice.
- No hype words: revolutionary, seamless, supercharge, unlock.
- Explain the benefit before the feature name.
- Release notes: one line per change, grouped by Added / Fixed.

## Example
Before: "Unlock seamless collaboration with our revolutionary
sharing feature."
After: "You can now share a dashboard with a link. Viewers
don't need an account."

The description line matters most: it is what tells Claude when the skill applies. Write it as a trigger ("Use when..."), not a label. For building and packaging custom skills, see how to create a Claude skill ZIP.

When not to make a skill: if a preference applies to every single response, it belongs in account instructions. If it applies only to one project, project instructions are simpler.

A 20-Minute Starter Setup

If you are starting from nothing, or cleaning up an old setup, this order gets you to a working configuration without overbuilding:

1. Audit what is already there (3 minutes).Open Settings > Your Claude and read your current instructions. Check Customize > Skills for any migrated styles; Anthropic has said migrated styles may start disabled, so enable only the ones you still want.
2. Rewrite account instructions (5 minutes). Cut them to the lines that are true for every chat. Five to ten bullets is plenty.
3. Create one project for your main body of work (5 minutes). Write project instructions for audience and goals, and add two or three reference files.
4. Pick one repeatable task for a skill (5 minutes). A writing voice, a report format, or a review checklist. Write the trigger description first.
5. Review memory (2 minutes). Look at what Claude has picked up about you and correct anything outdated.

Then stop and use it for a week. The biggest failure mode with personalization is building elaborate rules up front for situations that never come up. Add rules when you notice yourself correcting the same thing twice, and remove rules when you notice Claude follows the behavior without them.

A quick way to test a change: ask the same three representative questions before and after (one quick factual question, one piece of writing, one analysis). If the quick question gets worse after an edit, the rule probably belongs in a project or skill rather than at the account level.

Troubleshooting: When Claude Ignores Your Setup

• Conflicting layers. Account instructions say "be concise", project instructions say "be thorough". Remove the conflict instead of hoping Claude picks the right one.
• Vague rules. "Write well" does nothing. "No paragraph over four sentences" is checkable.
• Too many rules. Thirty rules dilute each other. Keep the ones that change output.
• Long conversations. In very long threads, early instructions can lose weight against recent context. A fresh chat often restores adherence. See why Claude loses track in long chats.
• Memory disagreeing with you. If Claude has picked up an outdated fact, review memory under Settings and correct it.
• Skill not triggering. Rewrite the description so it names the tasks where it should apply, and confirm code execution is enabled.

ClaudeKit Workflow Note

Instructions and skills shape how Claude responds. ClaudeKit, a free browser extension for claude.ai, handles the parts around that. Its prompt library lets you save prompts, including from any Claude message, organize them by category, and reuse them with one click, which is a good home for task prompts that are too specific for account instructions but not worth a full skill.

The live token counter shows a word count and token estimate while you type (estimates, roughly within 10%), which is a quick way to see how much a long pasted instruction block or reference adds to a message. ClaudeKit stores prompts locally in your browser and does not require an account.