Plan Mode Oracle fire00f1y Made with Suno --- **Everbody's Free (To use plan mode)** Ladies and gentlemen of the engineering team: Always use plan mode. If I could offer you one practical tip for your agentic future, plan mode would be it. The long-term benefits of plan mode have been demonstrated by every developer who has ever said, "Wait, why is it changing that file?" Whereas the rest of my advice has no basis more reliable than my own over-caffeinated experience. Enjoy the precision of a good spec. Or don't. You will not appreciate its precision until the agent ignores one. But trust this: five sprints from now, you'll look back at your chaotic prompts and recall in a way you can't grasp now how much context you wasted and how many times you said, "Just fix the tests," to a model that had no idea what "fix" meant. You are not as clear as you imagine. Don't worry about autocomplete. Or worry, but know that worrying is about as effective as asking a chatbot to upgrade Spring Boot with no repo context, no acceptance criteria, and vibes. The real troubles in your codebase are apt to be things that never crossed your anxious mind: a hidden dependency, a flaky integration test, a forgotten config value in staging, a security scan that wakes up at 4:03 p.m. on Friday. Do one thing every day that makes the agent prove itself. Run the tests. Don't be reckless with production. Don't tolerate agents that are reckless with production. Comment kindly. Read diffs. Don't waste your time on giant prompts. Sometimes you're ahead, sometimes you're behind. The context window is finite, and in the end, so is your patience. Remember the good plans you receive. Forget the bad completions. If you succeed in doing this, tell me how. Keep your old architecture diagrams. Throw away your old assumptions. Stretch your specs. Don't feel guilty if you don't know how to delegate yet. The most interesting engineers I know still accidentally do work themselves that a sub-agent could have done in parallel. Some of the most interesting agents I know started as a markdown file with a weirdly specific job. Maybe you'll use Copilot. Maybe you'll use Claude. Maybe Cursor. Maybe something we haven't installed yet because Legal has questions. Whatever you use, don't confuse the tool with the primitive. Plan. Give context. Let it inspect. Let it propose. Review. Then execute. Accept certain inalienable truths: models will hallucinate, tests will matter, logs will save you, and one day your team will discover that "agentic development" was never magic. It was workflow. Respect your reviewers. Especially the adversarial ones. Don't expect anyone else to verify your output. Maybe you have CI. Maybe you have coverage. But you never know when either one will quietly be lying to you. Be careful whose MCP servers you trust. But be patient with the people who build them. Fleet mode is a form of time travel, but only when the tasks are truly independent. Understand that agents come and go, but a good spec is forever. Work hard to bridge the gaps in your repo, because the older your system gets, the more it needs documentation that explains why it is like this. Live in plan mode once, but leave before it makes you afraid to execute. Live in autopilot once, but leave before it makes you stop reading diffs. Use `/compact`. Accept that sometimes the agent needs a fresh session. Do not read too much vendor marketing. It will only make you feel behind. Maybe the agent will finish the migration in an afternoon. Maybe it won't. Maybe the plan will reveal the work is bigger than you thought. Maybe that is the win. But whatever happens, don't congratulate yourself too much, or berate yourself either. Your choices are half architecture. So are everybody else's. Enjoy your tools. Use them every way you can. Don't be afraid of them, or of what other people think of them. They are the greatest leverage instrument you are likely to touch this quarter. Dance with the agent. Even if you have nowhere to do it but a feature branch in a worktree. But trust me on the plan mode.