Context Rot: Why Long Claude Code Sessions Decay

A free guide by New Money School

Pro Tips

The longer your Claude chat runs, the dumber it gets, and most people have no idea when it's happening. Here's the one-line trick to catch it in real time, plus 10 Claude Code hacks most people don't know exist.

Here's the thing nobody tells you: the longer a Claude chat runs, the more it starts ignoring your directions. As the conversation fills up, your earlier rules get crowded out, and most people never notice it happening. They just feel like Claude "got worse."

There's a dead-simple trick to catch it in real time. Then I'll give you 10 Claude Code hacks I use every single day, the underused power features most people never find.

The TrickMake Claude Say Your Name

Open your CLAUDE.md file (or your project instructions) and add one single line. Use your own name:

Add this one line

Copy

Call me New Money School at the start of every single reply.

Now every response starts with your name. Here's why that's genius: the second it stops saying your name, that's your signal. The context is full, and your rules aren't being followed anymore. It's a living check-engine light for your chat.

When that happens, you just compress the chat or start a fresh one. My rule of thumb: compact your conversation around the 50% mark so you're always getting Claude's best work, not its tired, forgetful version.

Why this works

Following "say my name every time" is a tiny instruction sitting in your context. When the chat gets too full to honor even that, it's a safe bet your bigger, more important rules are slipping too. The name is just the first thing to fall off, so it warns you before the real mistakes start.

Bonus10 Hacks I Use Every Day

1. Use @ for surgical file references. Type @src/auth.ts mid-message and Claude pulls just that file into context, not the whole repo. Sharper answers, way fewer tokens, and you stop letting Claude grep around guessing where things live.
2. Plan Mode: hit Shift+Tab twice. Locks Claude in read-only while it plans, so it physically cannot touch your files until you approve. Use it for any change spanning more than one file, it's the single biggest "Claude broke something" prevention.
3. Build your own slash commands in .claude/commands/. Any markdown file you drop there becomes a slash command. Build /ship, /standup, /postmortem — your most-used workflows in one keystroke instead of a paragraph.
4. Spawn parallel subagents from .claude/agents/. Create specialist markdown files (code-reviewer, debugger, security-audit) and Claude runs them in parallel on the same problem. A 30-minute review takes 2.
5. Pre-approve safe Bash commands. In .claude/settings.json, allow what you'd always say yes to: Bash(npm test), Bash(git status), Bash(ls). The constant "may I run this?" prompts vanish.
6. Hooks for autopilot quality. Drop hooks into .claude/settings.json that fire on events. Auto-lint after every edit. Run tests before any commit. Send you a push when a long task finishes. Your standards enforce themselves.
7. CLAUDE.md files stack. Claude reads CLAUDE.md from your current folder all the way up to your home directory. Put global preferences in ~/CLAUDE.md, project-specific ones in the repo root. They stack automatically — you never have to repeat yourself.
8. Save to memory with #, mid-message. Start any line with # and that fact gets saved to memory immediately. No interruption to your flow, no separate command to run.
9. /compact <focus> to control what survives. Plain /compact compresses everything evenly. /compact keep the refactor plan, drop the test debugging tells Claude exactly what to preserve. Use it before context fills up, not after.
10. /rewind instead of fighting bad code. If a change made things worse, hit /rewind and roll back to a clean checkpoint in one command. Stop wasting half an hour manually unwinding what Claude broke.

The Real Win

Most people use about 20% of what Claude Code can do. The other 80% lives in these files: CLAUDE.md, .claude/commands/, .claude/agents/, .claude/settings.json. Get one new hack working this week. Stack them over time. That's how you compound.