10 More Claude Code Tricks
You already got trick #0: name yourself in CLAUDE.md, watch when Claude stops using it, that's your context-full alarm. Reset at 50%, not 90%.
Here's 10 more. No docs regurgitation. These are the moves I actually run daily.
Kills: repeating the same coding standards in every new project.
~/.claude/CLAUDE.md
— put your standing rules here (style, commit format, tone). Project-level CLAUDE.md only holds what's unique to that repo.
The global file loads in every session automatically. You stop re-teaching Claude who you are on day one of every repo.
Kills: pasting the same multi-paragraph review prompt for the 50th time.
mkdir -p .claude/commands && nano .claude/commands/review.md
— write your prompt, use $ARGUMENTS for inputs, invoke with /review.
A paragraph becomes 7 keystrokes. It's version controlled with your repo, so your whole team gets it free.
Kills: dragging dead context forward when you switch tasks entirely.
/compact
mid-task, work isn't done.
/clear
when you're moving to a different task, not just a smaller version of this one.
Compact preserves continuity but it's a compression, not a clean slate. Clear gives Claude zero ambiguity about what's relevant right now.
Kills: Claude charging into a refactor on a wrong assumption you only catch after files are already edited.
shift+tab
toggles plan mode before a big change. Claude proposes the approach, you approve before it edits a single file.
Cheap to read a plan. Expensive to revert a bad refactor.
Kills: catching an unlinted file in CI instead of before it ever got committed.
// .claude/settings.json
"PostToolUse": [{ "matcher": "Edit|Write",
"hooks": [{ "type": "command", "command": "npx eslint --fix $CLAUDE_FILE_PATH" }] }]
This is the only trick on the list that doesn't rely on the model remembering. It's a shell command, not a request. It runs every time, full stop.
Kills: typing out a paraphrase of an error log that already exists on disk.
cat error.log | claude
or reference
@filename
directly in your prompt.
Claude reads the actual file. Not your summary of it, which is where the bug you're missing usually hides.
Kills: closing your laptop mid-task and re-explaining the entire architecture tomorrow.
claude -c
resumes your last session in this directory.
claude -r
opens a picker if you've got more than one going.
Full message history and tool results restored. Zero re-explaining.
Kills: describing a broken UI in words and watching Claude fix the wrong thing.
It sees what's actually broken instead of guessing from your description of what's broken.
Kills: Claude speculating about your database schema instead of just querying it.
mcpServers
in settings.json.
Turns Claude from "writes code about your system" into "actually touches your system."
Kills: writing tests, updating docs, and researching a library one after another, burning the same context window for all three.
Parallel work, no context bleed between unrelated tasks.
My #1 is trick 5. Everything else on this list relies on you remembering to do it. Hooks don't. That's the difference between a good habit and a guaranteed one.
Reply and tell me which trick you're stealing first, or what you want broken down next.
Next reel: the exact 3 MCP servers I run every single day, and what each one is actually for.
— Mike
@mikemeansbusiness_ai