Power-user tips for Cursor, Claude Code and Codex

Power-user tips for Cursor, Claude Code and Codex are verified habits that remove friction from agent sessions: raising effort before switching model, rewinding instead of patching forward, parallel sessions in Git worktrees, and headless runs with structured output. Each tip gets one line, checked 2026-09-26 (Cursor features 2026-08-28), and a link to its canonical guide.

Daily users still leak time: a session drifts after two hours, a review is re-run by hand, a second task waits for the first. Each tip removes one leak. Adopt one or two a week: /effort and /rewind first, then worktrees, then budget-capped headless runs.

Which tips work the same way in all three tools?

Four habits exist in all three tools under different names.

Habit	Claude Code	Codex	Cursor	Canonical page
Plan before a multi-file change	/plan	/plan	Plan Mode	Plan mode before code
Run a second task in its own checkout	claude --worktree feature-a	codex --worktree	Worktrees	Parallel agents
Set a finish line the agent checks before stopping	/goal	/goal	/goal (rolling out since 2026-08-19; checked 2026-08-28)	The /goal command across tools
Get an automated review	/code-review	codex review --uncommitted	Bugbot (runs once the PR is open)	Agent PR review

Tips for each tool

Twenty-six tips, with exact commands for Claude Code 2.1.283 and Codex CLI 0.157.1 and Cursor feature names, re-checked monthly (next review: October 2026).

Claude Code

Checked against 2.1.283 (latest); stable was on 2.1.274 on 2026-09-26, so a flag can be missing there.

Tip	Command	Where it is explained
Raise effort before you change model	/effort high	Features you haven’t tried
Rewind to a checkpoint instead of patching forward	Esc Esc, or /rewind	Features you haven’t tried
Ask a side question mid-run	/btw what does this flag do?	Features you haven’t tried
Run a shell command without a reasoning turn	!npm test	Features you haven’t tried
See what fills a fresh session’s context	/context	Cost control
Clear between unrelated tasks, compact with a focus	/clear, /compact keep the migration plan	Efficiency hacks
Reopen the session linked to a pull request	claude --from-pr 1234	Efficiency hacks
Hand a task to a background session	claude --bg "fix the flaky auth test", then claude attach <id>	Agent view
Cap spend on a scripted run	claude -p --max-budget-usd 5 --output-format json "…"	Cost control
Review your own diff before you push	/code-review	Review automation

Codex

Checked against Codex CLI 0.157.1.

Tip	Command	Where it is explained
Attach a file or run a shell command	@src/auth.ts, !pytest -q	Efficiency hacks
Resume or fork the last session	codex resume --last, codex fork --last	Efficiency hacks
Switch settings per job with a named profile (create ~/.codex/review.config.toml first)	codex --profile review	Efficiency hacks
Review local changes, or a branch against main	codex review --uncommitted, codex review --base main	Code review
Start a parallel session in its own worktree	codex --worktree	Worktrees
Get fields a script can parse	codex exec --output-schema schema.json -o result.json "…"	Non-interactive mode
Use a permission profile instead of --sandbox (beta)	codex -c default_permissions=":workspace"	App mastery
Set a finish line for a long-running task	/goal	The /goal command

Cursor

Verified on cursor.com on 2026-08-28; menus and key bindings change often, so this tab names features only.

Tip	Feature	Where it is explained
Plan before a multi-file change	Plan Mode	Agent modes deep dive
Put conventions in a rule instead of repeating them	Rules	Project rules
Run agents in parallel in isolated checkouts	Worktrees	Agents Window
Repeat a prompt until the tests pass, with a stop condition	/loop bundled skill	Test-driven development in Cursor
Move a long task off your machine	Cloud Agents	Cloud Agents and Automations
Start agents from Slack, Linear, Sentry or a schedule	Automations	Cloud Agents and Automations
Format and test after every agent edit	Hooks	Automation workflows
Get a bug review on every pull request	Bugbot	Bugbot

More Cursor shortcuts: Cursor time-savers.

Copy-paste prompt: find the tips your setup is missing

Audit this repository's agent setup. Read CLAUDE.md or AGENTS.md, the project
settings, MCP servers and skills. Report as a table: what loads into every
session and its rough context cost, which verification commands (test, lint,
typecheck) are documented, and whether a plan-first rule, a worktree setup
script and a pre-push review step are missing. Change no files. Suggest at
most three changes, most valuable first.

Copy-paste prompt: hand off before you clear or compact

Before I clear this session, write docs/handoff.md: the goal, decisions and
why, files changed, the commands that prove the current state (tests, lint,
typecheck) with their real output from running them now, and the next three
steps. Keep it under 60 lines.

How do you know a tip helped rather than just felt faster?

A tip changes how you drive the agent, never what counts as done: tests, lint and type checks still run, and a reviewer signs off on the evidence bundle.

Compare a week before and after: /usage (spend in Claude Code, limits in Codex) and how often a change passes its gates first time. If that rate drops, drop the tip.

When power-user tricks backfire

• A flag is “unknown”. Your install is older than this page; check --version and update.
• A parallel worktree fails its tests. New worktrees lack dependencies and env files; add the setup step from Parallel agents.
• /compact or /clear lost a decision. Run the hand-off prompt before either.
• Scripted runs overspend. Set --max-budget-usd on every claude -p run.
• A permission profile goes too far. Keep :workspace for trusted repositories; run untrusted code in a sandbox.

Where to go next after these tips

Quick wins in your first 24 hours The basics these tips assume.

Developer track The ordered reading path for developers.

Features you haven't tried in Claude Code Underused Claude Code features, setup to headless.

Reading evidence instead of code Why speed pays off only with solid verification.