The 4 Context Antipatterns
Claude Code is powerful. However, it is easy to accidentally mess up your own session if you're not careful with context. Anthropic’s Claude Code docs list five common failure patterns. I'm going to go over four of them.
Antipattern 1: The Overstuffed Fridge
You want to build a button. You start Claude Code. Halfway through, you ask Claude something random: "Hey, can you explain OAuth?" Claude answers. You go back to building a button. That's cool, but now your context contains both.
Claude's context is like a fridge. It has finite space. When you mix unrelated stuff in, everything gets harder to find. Quality degrades. Claude re-reads all that OAuth stuff every time you ask about the button. It is not giving 100% attention to the button anymore.
Instead, /clear (aliases: /new or /reset) between unrelated tasks. Doing so wipes context and starts you fresh. Your next task gets Claude's full attention.
If your mind wanders like mine, keep a separate Claude session just for those questions.
You can use /btw for a quick side question. The answer never enters your conversation, so your context stays clean.
A fresh context with a focused prompt is better than a cluttered one.
Antipattern 2: The Correction Rabbit Hole
Claude writes code. The result doesn't look right. You correct it: "Make the button smaller." Good, but something is still off. "Make the edges rounder." Again. Again. By the tenth correction, your context is full of attempts. Claude drowns in noise. Each correction doesn't erase the last one. It adds to the pile. "Instead, can you..." adds a new line. Each line is context spent.
Instead, after two failed attempts, /clear and restart with a better prompt. You want Claude to stay sharp. Don't clutter context with corrections. Make every token count.
Don’t do the tenth “no, try this.” Start fresh instead. This time, be precise: “In SubmitButton.tsx, reduce the button height from 48px to 40px, reduce the horizontal padding from 24px to 16px. Shift it 10px left by changing margin-left from 24px to 14px. Make the corners fully rounded with border-radius: 9999px. Keep the font size.” Include what you learned. A clean slate with a better prompt is better than noisy, endless corrections.
Catch mistakes early. Press Esc to stop Claude mid-action. If a turn already went wrong, press Esc Esc (or /rewind) and roll back to before it. A rewind removes the bad attempt instead of piling a correction on top.
Antipattern 3: The Marco Polo Session
You ask Claude: "Investigate why user authentication is slow." Claude reads files. More files. Then more files. Giant log files! Suddenly your context is thousands of lines deep. Claude explores things that probably don't matter.
Claude is like Marco Polo. It likes to explore, and when it does, it explores wide. Claude can read a lot of files faster than a human, but it also fills context with noise faster than I can fill my shopping cart on Black Friday with stuff I don't need. By the time you get an answer, your context is filled up with expedition notes.
Scope the investigation. Or hand it off. Instead of "Investigate why user authentication is slow," say: "Look only at the session lookup in src/auth/. Find slow database queries." Claude now has boundaries and it explores narrow.
Maybe you don't know the boundaries. If you want it to search far and wide, use a subagent. Ask: “Use a subagent to investigate why the login endpoint is slow. Read the relevant files and report back a summary.” The subagent explores in its own context. You get a clean summary. All the file-reading stays in the agent's context, not in your main conversation.
To spawn an investigation in the background while you work on the main task, use /subtask (requires Claude Code v2.1.212 or later). /subtask investigate why user authentication is slow forks a subagent that inherits your current conversation and reports back when it’s done. A subagent is like delegating to an assistant. "Go figure this out. Come back with a summary." The assistant can explore far and wide and that context stays with it. It never enters the main conversation. What comes back is just the brief result.
Antipattern 4: The Overstocked Staples
If Claude's context is your fridge, CLAUDE.md is your staples: milk, eggs, butter (almost every fridge has them, and almost every meal starts from them). Claude loads CLAUDE.md into every session, so whatever you put there takes up space in every conversation.
The problem starts when you have too many “staple” items. Do you really need 2% milk, 1% milk, and whole milk? Maybe just pick one. Do you need four half-empty bottles of different hot sauces instead of one? Do you need a dozen regular-sized eggs and a dozen large cage-free eggs?
Pretty soon you can't find your hot sauce and your milk spoils. A bloated CLAUDE.md works similarly. Important rules get lost in the noise. Claude starts ignoring half of them. Like that milk, old rules go stale. Anthropic suggests keeping each CLAUDE.md under 200 lines. For each line, ask: “Would removing this cause Claude to make mistakes?” If Claude does fine without it, cut it.
Keep in mind that splitting CLAUDE.md into @imports doesn’t save space. Imported files load at launch too.
Instead, turn knowledge or workflows you only need sometimes into a skill, which Claude loads only when it’s relevant. Rules for one part of the codebase belong in a path-scoped rule, which loads only when Claude touches matching files. A rule that must happen every single time belongs in a hook.
The Pattern
These antipatterns share one theme: context pollution. Claude's context is filled with unrelated tasks, failed corrections, endless exploration, and instructions it doesn't need. Claude's output degrades not because Claude got dumb, but because it’s wading through noise and has less room to think.
Treat these as guardrails. Use them so Claude doesn't get sloppy. Context is your most limited resource. Don't pollute it.