A developer reviews code and system status on a large monitor at a tidy home office desk.
There's a stage in every power user's life when the setup starts to look like a gaming PC with 14 RGB utilities in the system tray. Everything was installed for a reason, and nobody can remember what the reason was. XDA writer Mahnoor Faisal has hit that stage with Claude Code, Anthropic's agentic coding tool. Her conclusion will make a lot of developers wince: several of her "advanced" habits were making the tool worse.

She wrote the piece as a first-person audit. She stripped her heavily customized setup back and compared it with something closer to what a beginner would run. To be clear about what it is: this is an experience report, not a benchmark. There are no token counts, timed runs or controlled tests. Still, Anthropic's own documentation backs up most of her conclusions, and it adds some specific numbers and commands that make the lessons useful for anyone running Claude Code on Windows, in WSL or anywhere else.

The four habits that backfired​

Faisal names four kinds of accumulated clutter:

  • Too many MCP servers. She had connected everything: community-built servers for NotebookLM, connections to Adobe, Canva, Notion, Slack and Gmail, the usual developer servers for GitHub, databases, Vercel and Playwright, and even meeting tools like Read AI.
  • Memory layered on memory. She ran add-ons called claude-mem and claude-memory on top of Claude Code's built-in memory features.
  • An overstuffed CLAUDE.md. Coding conventions, workflow preferences, reminders, edge cases and every mistake Claude had made once all went into the file that loads with the project.
  • Plan mode by reflex. She switched into Plan mode before almost every prompt, even for edits that take two minutes.

Her lesson isn't "features are bad." It's that adding things to your setup isn't the same as improving it. Each extra should earn its place by solving a problem that keeps coming back.

Section summary: This is a personal audit of four kinds of clutter. The anecdotes are hers. The context below comes from Anthropic's documentation.

MCP sprawl: less of a context hog, still a choice problem​

Faisal's first complaint is about choice. Every server adds tools Claude might have to pick from, and overlapping tools can make that selection messier. She also points out that before Claude Code's Tool Search started deferring tool definitions, a pile of servers could use up a large share of context before you typed anything.

That second point is mostly history now, and the timing matters. When Anthropic launched Tool Search, VentureBeat described it as bringing "lazy loading" for AI tools, allowing agents to dynamically fetch tool definitions only when necessary. In that first version, the system automatically detects when tool descriptions would consume more than 10% of the available context.

Anthropic's current feature overview goes further. Tool Search is on by default. At session start, Claude Code loads tool names and server instructions, and the full JSON schemas wait until Claude actually needs a specific tool. Idle MCP tools therefore cost very little context. That's not zero, but it's far from the old problem of losing half your context window at startup.

There's a catch for enterprise users. A community-maintained Claude Code guide on GitHub notes that Claude Code turns it off when ANTHROPIC_BASE_URL points to a non-first-party host, since most proxies don't forward the blocks it needs. If your organization sends Claude traffic through a corporate gateway or proxy, the old bloat problem may still apply to you. Check your own session rather than assuming the default behavior. The same guide lists an ENABLE_TOOL_SEARCH setting with a threshold mode, in which tools load upfront until their definitions reach 10% of the context window, then all are deferred.

Deferred loading also doesn't fix the choice problem. Anthropic's API documentation for its tool search feature suggests using it when tool selection accuracy drops as your toolset grows. Lazy loading makes a big toolbox cheaper to carry. It doesn't make it tidier.

How to audit your MCP servers:

  1. Run /mcp to see every connected server and its connection status.
  2. Run /context all to see how many tokens each loaded MCP tool is using.
  3. Disconnect servers you don't use regularly, especially ones that overlap, such as two servers that can both reach the same service.
  4. Keep the integrations that give Claude data or actions it can't get any other way.
  5. Where a command-line tool will do, consider using it instead. Anthropic's best-practices guide calls CLI tools like gh, aws and gcloud the most context-efficient way to work with external services.

Section summary: Tool Search has largely fixed the startup context cost, unless a proxy disables it. The number of overlapping tools can still muddy Claude's choices, so prune based on actual use.

Memory add-ons: when remembering hurts​

Faisal's reasoning here holds up: a memory system can keep something that was true three sessions ago but isn't anymore. Claude then has to reconcile what it remembers with what's actually in the repository, and stale context can push it in the wrong direction. She also found that tasks in the same directory often had little to do with each other. Sharing a project didn't mean one session's context would help the next.

A caveat applies. Her piece doesn't describe how those plugins were configured or what they loaded, so this isn't a verdict on claude-mem or claude-memory in general. The safer takeaway is that anything persistent can go stale, and stale context is worse than none.

Anthropic's best-practices guide lines up with her instinct. It recommends running /clear between unrelated tasks and warns that long sessions full of irrelevant context can hurt performance. If you've corrected Claude more than twice on the same problem, Anthropic suggests clearing and starting over with a better prompt. A fresh session with a sharper prompt usually beats a long one full of accumulated corrections.

Section summary: Memory is only helpful while it's accurate. Check stored context against the live repository, and don't be afraid to start with a clean slate.

CLAUDE.md: the one file that costs you every time​

This is where Faisal's experience and Anthropic's guidance agree most closely. She found that as her CLAUDE.md grew, important rules got buried, old guidance outlived its usefulness, and instructions meant for one kind of work leaked into unrelated tasks.

Anthropic now gives a concrete limit. Its feature overview says to keep CLAUDE.md under 200 lines. Move reference material to skills, which load on demand. To get trim proposals for a checked-in CLAUDE.md, run /doctor. The best-practices guide puts it even more bluntly: bloated CLAUDE.md files make Claude ignore your actual instructions. It also offers a simple test for each line: would deleting it make Claude make mistakes? If not, cut it.

The documentation also explains where things should go instead:

Put it hereWhenExamples
CLAUDE.mdClaude needs it in every sessionBuild commands, core conventions, "never do X" rules
.claude/rules/Guidance for specific languages or foldersRules scoped to certain file paths
SkillsMaterial needed only sometimesAPI docs, style guides, deploy checklists
HooksAnything that must happen every timeRunning ESLint after every edit, blocking changes to sensitive files

Anthropic's documentation makes a point about hooks that matters for teams: an instruction such as "never edit .env" in CLAUDE.md is a request, not a guarantee. If a rule really has to be enforced, a PreToolUse hook that blocks the action is the real guardrail. Some of Faisal's "things Claude got wrong once" entries probably belonged in a hook or nowhere, not in an ever-growing text file.

Section summary: Keep CLAUDE.md under about 200 lines and limited to rules that apply to the whole project. Move reference material into skills and enforcement into hooks.

Plan mode: tapping the brakes at a red light that isn't there​

Faisal likes Plan mode, and so does Anthropic. It lets Claude read files and run commands to explore, then propose an approach without touching source code. Her problem was using it by default, which added a round of planning for tasks that needed only a couple of edits.

Anthropic's best-practices guide backs her up almost word for word. It says Plan mode is useful but adds overhead. For small, clearly scoped fixes like a typo, an extra log line or a renamed variable, it suggests just asking Claude to do the job. Its rule of thumb: if you could describe the diff in one sentence, skip the plan. Planning pays off when you're unsure of the approach, when the change touches several files, or when you don't know the code well.

In the terminal, you enter Plan mode by pressing Shift+Tab until the status bar shows plan mode on, or by starting with claude --permission-mode plan. In the Desktop app's Code tab, Anthropic's documentation says Plan applies only to the current session, and after you approve a plan you switch to Accept edits or Manual to carry it out. Mode names have changed between versions, so what you see may differ.

Section summary: Use Plan mode in proportion to how complex and risky the task is. The one-sentence-diff test is a good filter.

The bigger picture​

Why does this happen to experienced users? Because configuration feels like progress. Every "you're using it wrong" post offers a small upgrade, and trying one costs a few seconds, so the clutter builds up without anyone noticing.

Anthropic's feature overview actually recommends the opposite: start with CLAUDE.md and add extensions only when a specific need shows up. Faisal's audit makes the same case from the user's side. The best setup is often the one that gives Claude exactly what it needs and nothing more.

One caution: this doesn't show that beginner setups beat advanced ones. A well-chosen database server, a skill for your deploy process or a guardrail hook can clearly be worth it. The goal is to audit what you've added, not to throw it all out.

Your quick audit checklist:

  • /mcp and /context all: remove unused or overlapping servers.
  • /doctor: get suggested cuts for a checked-in CLAUDE.md, and aim for fewer than 200 lines.
  • /clear between unrelated tasks, and check persistent memory against the repository.
  • Skip Plan mode when the change fits in one sentence.
  • Move rules that must be enforced into hooks instead of prose.

If your own Claude Code setup has grown more plugins than your browser has extensions, now's a good time to trim it.

 

References

  1. I compared my Claude Code workflow to a beginner's, and some of my habits were making things worse XDA 2026-09-30T20:00:20+00:00
  2. Perluas Claude Code - Claude Code Docs code.claude.com
  3. Best practices for Claude Code - Claude Code Docs anthropic.com