For long Codex work, write the plan down first. A good PLANS.md has milestones, validation commands, and a decision log. The agent gets a map. You get resumable progress instead of a mystery thread full of half-remembered intent.
Tips
They almost feel like cheating.
Small tricks that save time, clicks, and frustration. Borrowed, tweaked, and battle-tested.
AGENTS.md is Codex context that survives the chat. Put the repo's test command, style rules, risky files, and review expectations there. Keep it short. If you would repeat it in every prompt, it belongs in the file.
Close the PR loop inside the agent. After it opens a pull request, have it read the review comments, inspect failing checks, patch only the requested scope, and push again. The prompt is boring: fix the PR feedback and rerun the checks. Boring is the point.
git worktree add ../repo-auth -b agent/auth. Now Claude, Codex, and you can all work the same repo without stomping each other. One branch per agent, one terminal per branch, normal PR review at the end.
Branch the conversation before the risky idea: claude --continue --fork-session. In the fork, double-tap Esc and rewind to the decision point. Let one thread chase the refactor while another keeps the boring fix alive.
claude --chrome turns frontend debugging into a real browser loop. Point it at localhost, let it read console errors and DOM state, then fix the code from the same session. Use Playwright when you need repeatable tests. Use Chrome when you need to see what is actually broken.
Double-tap Esc -> rewind menu -> "Summarise from here." Compresses all the dead-end trial-and-error into a dense summary while preserving everything before it. You keep the lessons, ditch the noise, reclaim your token budget.
--worktree gives each agent its own sandbox. Two agents, same repo, zero collisions. Each gets an isolated working tree. When they're done, merge normally. claude --worktree feature/auth-refactor. claude --worktree feature/dashboard-ui.
Claude Code adds its name to your commits by default. To turn that off: Add this to ~/.claude/settings.json.
Ask your agent "what's the dumbest way to solve this?" Sometimes that's the right answer.
--fork-session is git branch for your context window. Load a master session with your architecture, docs, and standards. Fork it for each feature. Clean, independent context every time - no rebuilding, no bleed-over.
Ctrl+G opens your $EDITOR mid-prompt. Write complex prompts in Vim, Neovim, VS Code - whatever you've configured. Full macros, multi-line editing, syntax highlighting. Save and quit, it flushes straight into Claude. Prompt quality goes up when you can actually see what you're writing.
If you open a PR during a Claude session with gh pr create, it links the session ID automatically. Later, claude --from-pr 123 rewakes that exact agent - full context, original tradeoffs intact.
One task. One agent. One problem. Everyone spins up OpenClaw and dumps every skill, tool, and agent into the mix at once. It gets confused, you get chaos. Start small. Verify. Then stack.
Open a terminal. Point Claude Code at your OpenClaw machine. Let them talk. One builds. One questions. They figure it out together so you don't have to.
When you're directing Codex or Claude Code, use ⌘ + Ctrl + Shift + 4, then ⌘V. Draw a box around exactly what's broken and paste it straight into the chat. No screenshots folder, no dragging files - just grab and drop. Error messages, UI bugs, terminal output, whatever. Show, don't tell.