← All posts

25 Claude Habits That Separate the Top 1% From Everyone Else

Memory, models, prompts, deliverables and privacy: 25 small changes that make Claude noticeably more useful and keep your limit intact.

November 16, 2026 11 min read Cost Optimization
TL;DR

Work in the desktop app, default to Opus 5.5 at High effort, keep memory on but short, and clear out instructions written for older models. Give Claude one goal with a clear finish line, edit your prompt instead of arguing, and ask for the finished thing (a deck, an HTML page, a dashboard), not a draft. Below are all 25 habits in five groups.

Where This List Comes From

This guide is an English adaptation of a Russian-language playbook, "25 привычек в Claude" by the Telegram channel Силиконовый мешок, itself based on a November 2026 discussion on r/ThinkingDeeplyAI. The playbook's author checked the numbers and setting names against Claude's help center and an Anthropic post on memory as of November 2026, and corrected the original where it was wrong. Menus and toggles move often, so if a setting below has a different label in your app, search settings for the keyword.

Part 1: Set Claude Up for Yourself (Habits 1–8)

1. Claude Cowork is gone. There is just Claude now

Cowork was merged into Claude in September, so there is no separate mode any more. Use the desktop app and forget about Claude in the browser. The desktop app is the one that works directly with folders on your disk, so you don't have to upload files into a chat. You'll also open Claude Code more often, because it lives there too.

2. Default to Opus 5.5 at High effort

For everyday work pick Opus 5.5, not Fable 5.2. Fable 5.2 is extremely capable, but it eats your limit noticeably faster: on Max only half of the weekly limit is allotted to it, and on Pro it is billed separately. Opus 5.5 at High (the Effort setting next to the model name) handles the work and lasts the whole day. Whatever you choose, don't change the model in the middle of a chat: the new model reads the whole history from scratch, which is a needless drain on your limit.

3. Keep memory on. Claude works noticeably better with it

Anthropic compared models with and without memory. In CL-Bench, a test where the model keeps notes from session to session, Opus 5.5 with memory scored 28% higher. Memory is on by default in chat. Check the toggle in Settings, under Your Claude.

4. But keep memory short

Before one cleanup, Claude remembered 105 facts about the author. After a single prompt, 26 remained; the rest were outdated or wrong. Ask Claude to clean up its memory of you and keep only what is still useful. What it remembers is visible in Your Claude under Manage.

Review everything you remember about me. Remove anything outdated,
duplicated or no longer true, and keep only what is still useful.
Show me what you removed.

5. Delete the "about me" file. Don't turn it into a skill

Many people turned their biography into a skill, an instruction Claude loads by itself so it "knows" them. It doesn't help. Memory already handles this, and handles it better, while a skill just slips the same stale paragraph into every task.

6. Clean up the instructions field

Open "Instructions for Claude" in settings and delete everything written for older models. The same goes for old skills. Long rule lists and ALL-CAPS commands only get in the way of newer models. Keep it short and specific, and let one clear goal per prompt do the rest.

7. Don't dump every file into the chat

Stop dropping a "context" folder into every chat. When Claude has 40 files at hand, it retells them instead of thinking. Give it the one file the task really needs and let it work out the rest.

8. One project per client or task. Turn off shared memory inside it

On the project page, switch off Use account memory. Then what Claude knows about client A won't surface in work for client B.

Part 2: Protect Your Chat and Your Limit (Habits 9–12)

9. One bad answer ruins a long chat

Once a wrong answer is in the chat, Claude builds on it, and every following answer comes out a little worse. Patching on top is pointless. Go back and remove it, as habit 10 explains.

10. Don't argue with Claude. Edit your prompt

Arguing burns tokens and you still won't finish the task. Click Edit on the message that preceded the bad answer, rewrite it and press Save. The bad answer stays in the old version of the chat and no longer affects the new one.

11. Long chats burn tokens. Summarize and move

With every new message Claude re-reads everything above it. When a chat has grown huge, ask:

Collect everything we decided and everything that is still open,
so I can paste it into a new chat.

Open a new chat with that text.

12. Check automatic model switching

In Settings, under Capabilities, there is a toggle called "Switch models when a message is flagged". If the safety filter considers a request to Fable risky, Opus will answer instead and the chat stays on Opus for good. Claude warns you, but in long work that is easy to miss. With the toggle off, the chat simply pauses, and you decide whether to rewrite the request or change the model.

Part 3: Frame the Task So It Gets Understood (Habits 13–20)

13. Don't give Claude steps. Give one concrete goal

Step-by-step instructions were for older models. New ones plan better than you will spell out. Say what the result should be and stay out of the way.

14. Tell Claude what "done" means

Define the finish line: "Done when you have a one-page document my manager can forward without edits." Now Claude knows what a good result looks like and when to stop.

15. Ask Claude to restate the goal before it starts

End your prompt with "Before you begin, restate my goal in your own words." If it understood wrongly, you find out immediately instead of after a whole wasted answer.

16. End every prompt with "ask me questions". Especially the first one

When Claude doesn't know something, it fills in the blanks. Let it ask instead: add "Ask me all the questions you need before you start." The questions themselves show what you forgot to say.

17. Don't hint at the answer. Ask for criticism

Ask "why is this a good idea?" and you get a list of reasons it's good. Ask differently: "What's the strongest argument against? Where does this break?" That is where the useful part is.

18. Ask for 20 very different options. Not five

Claude gives the most typical answer first; that's how it works. To reach the strange and non-obvious ones, use this prompt:

Give me 20 very different options. For each one, say how likely
another AI would be to produce it. The last two should be below 10%.

19. Talk, don't type. Then send a screenshot

Dictate by voice. You'll say three times more than you would type, and those details are exactly what makes an answer good. For what you see on screen, don't describe it in words; send a screenshot.

20. A GitHub link works almost like a skill

Found a repository that does what you need? Paste the link. Claude reads the public repository and builds in its image. Nothing to configure.

Part 4: Ask for the Finished Thing (Habits 21–23)

21. Don't ask for an outline. Ask for the presentation

Outline first, then edits, then slides: three rounds, and each one spends limit. Ask for the finished deck right away. Fixing finished slides is faster than negotiating a plan.

22. Ask for HTML, not a Word file

Add "make an HTML page" to the prompt and say who it is for. Claude will lay out a page with graphics and diagrams itself. It looks better and there is almost nothing to break.

23. Paste the numbers. Ask for a dashboard

Ask Claude to "analyze" a table and you get a paragraph of text. Paste the numbers, say "turn this into a dashboard", and add who will be looking at it. You get the point in 10 seconds.

Part 5: Don't Give Away More Than You Need To (Habits 24–25)

24. Don't press thumbs up or thumbs down in work chats

When you rate an answer, the whole conversation is kept by Anthropic for up to five years and may be used to train models. If the chat contains client work or anything personal, you don't want that. Skip the rating there.

25. Don't connect your bank. Export a spreadsheet

There is no need to give Claude direct access to your bank account. Export a CSV with the transactions you need from the bank, remove the account numbers and paste it into the chat. The analysis is the same and the risk is far smaller.

Where this overlaps with Claude Code: several habits map directly onto it: one clear goal and a definition of done (13, 14), asking questions first (16), summarizing and starting fresh instead of arguing (10, 11), and keeping always-loaded instructions such as CLAUDE.md short (6). See also our guide to 22 rules to stop hitting Claude token limits.

The One-Page Checklist

  1. Cowork is gone; use the desktop app
  2. Opus 5.5 at High by default; never switch models mid-chat
  3. Keep memory on
  4. Keep memory short
  5. Delete the "about me" skill
  6. Clean up the instructions field
  7. Give one file, not forty
  8. One project per client; turn off account memory inside it
  9. One bad answer spoils a long chat
  10. Edit your prompt instead of arguing
  11. Summarize long chats and move to a new one
  12. Check automatic model switching
  13. One concrete goal, not steps
  14. Define "done"
  15. Have Claude restate the goal first
  16. End with "ask me questions"
  17. Ask for criticism, not agreement
  18. Ask for 20 very different options
  19. Talk, then send a screenshot
  20. Paste a GitHub link
  21. Ask for the presentation, not the outline
  22. Ask for HTML, not Word
  23. Paste numbers, ask for a dashboard
  24. No thumbs up/down in work chats
  25. Export a CSV instead of connecting your bank

Each one takes a minute at most. Pick one to try first.

Every Claude Code command, flag and shortcut in one place.

Open the Claude Code Cheatsheet →