What’s in this guide?
The full list of the 10 mistakes almost everyone makes with Claude Code, and the fix for each one. Short and to the point.
Claude Code is a powerful tool. But there are mistake patterns almost everyone falls into when they start using it. I’ve tracked them for months and identified ten recurring ones. The list below is the first filter every new project of mine passes through.
Worth reading in one sitting, and revisiting once a month to make sure you’re not slipping back into the same patterns.
Source: MasteringAI by Mark Fershteyn. Translated, condensed, and adapted.
Problem: Every new conversation, you explain again what the project is, how it’s built, how you like to work. Another ten valuable minutes down the drain.
Fix: Create a file called CLAUDE.md in the project’s root folder. In it, write the technologies you use (Python, React, etc.), your coding rules, and the things not to do. Claude reads this file automatically at the start of every conversation, without you asking.
Problem: In a long conversation, Claude starts forgetting constraints, repeats suggestions you already rejected, or goes in circles. Instead of stopping, you keep pushing. (Context = all the information Claude carries through the conversation; the longer the conversation, the more overloaded the context, and the more Claude starts missing things.)
Fix: The command /clear starts a new conversation and resets Claude’s memory of the previous one. The moment you notice the signs, run it. Before clearing, save anything important (code, ideas, decisions) to files, because after /clear they’re gone from the conversation.
Problem: You start with “build this” without letting Claude plan first. Thirty minutes later you find out it built something that isn’t quite what you wanted, or built something good, but going in the wrong direction.
Fix: Before any significant task, enter Plan Mode. The shortcut: Shift+Tab twice. In this mode, Claude presents a detailed plan before writing a single line of code. You read, approve or revise, and only then it executes. Iron rule: never approve a plan you don’t fully understand.
Problem: After Claude planned and you approved, you execute in the same conversation. The hesitations and considerations from the planning phase sit on Claude and cause it to execute with a wobble.
Fix: Work in two separate conversations. First conversation, planning only. Save the approved plan. Open a new conversation (with /clear), paste in the plan, and ask Claude to execute. The execution comes out clean and focused.
Problem: “Analyze the data,” “build a website,” “fix the bug.” Generic instructions. Claude has to guess, and the guess comes from the average of the internet, not your specific needs. (A “prompt” = the written instruction you give Claude.)
Fix: A four-part formula that should appear in every meaningful instruction:
• Input: which files or data Claude should see.
• Action: what exactly to do (write, fix, analyze, create).
• Details: specific requirements and constraints.
• Format: what the output should look like at the end.
Problem: When working on large files, Claude “breaks” unrelated code in other places in the file, sometimes 2000 lines away from where you asked for a fix. A small change turns into a new bug you don’t notice.
Fix: Safe boundaries. Up to 1000 lines per file is ideal. Between 1000 and 2000, consider splitting. Over 2000, mandatory split before you ask for large changes. The split should be by role: one file for user login, one file for data management, one file for the interface, not an arbitrary cut by line count.
Problem: Long projects without stopping points. A hiccup at step 15, you start over from zero, because there’s no safe mid-point to go back to.
Fix: Any project that takes more than 30 minutes, ask Claude for a numbered step list at the start of the work. At key points, add [CHECKPOINT] markers. The marker tells Claude to stop there, summarize what’s been completed, and wait for approval before continuing. That way if something breaks at step 15, you go back to the previous checkpoint, not to the start of the project.
Problem: Every new conversation, you explain the same personal preferences: writing style, report format, standards you require.
Fix: Two tools. First, a global CLAUDE.md file (one file loaded across all projects) that lives at ~/.claude/CLAUDE.md. Second, Skills in the claude.ai interface: a collection of saved instructions you can call with a click, activating all of them together in any conversation.
Problem: Serial work, one after the other, on tasks that could run in parallel. For example: research on three different topics, or processing five separate files.
Fix: Sub-Agents are additional Claudes that the main Claude can run in parallel. Before starting a big task, identify parts that don’t depend on each other, and ask Claude “run a sub-agent for each of the following checks.” Instead of an hour of serial work, ten minutes in parallel.
Problem: “I’ve already invested 30 minutes in this conversation. If I just explain one more time…” And you keep pushing a conversation that’s already dying. The more time you’ve put in, the harder it gets emotionally to admit you need to start over.
Fix: Starting fresh is the professional move, not the surrender. The moment the thought “maybe I should start over” comes up, run /clear immediately, without negotiating with yourself. A new conversation with a clear plan will get out in 5 minutes what would have taken another hour to fix in the old one.
Fill out this short form and I’ll come back with a plan built around your business. Fill out the form →
Prefer to talk directly? Book a 30-min call →