Whole new projects get built in four sequential steps. Each step its own terminal. Knowledge travels between steps via Markdown files, not chats.
When something new is being made: a workflow that doesn't exist yet. An app. An integration. A bigger idea you haven't fully thought through. The flow forces you to think before coding, and to ask before thinking.
Each step produces a Markdown file. The next terminal reads that file and builds on it. Chats are tools, not knowledge stores. If you let the context window overflow, you lose substance. Instead: each step = new terminal, new chat, but the same Markdown truth.
New terminal with cr. You throw in the idea, Claude asks questions, clarifies options, writes along.
1. Stay in dialog until it clicks.
You keep the dialog going until the AI has no more follow-up questions and you have no more follow-up questions. The feeling must be: this step is really done.
2. Then send both instructions together in one message:
3. After that: /shutdown. Close this terminal, open a new one.
New terminal with cr. Paste the prompt from step 1. Claude goes deeper: constraints, architecture options, risks, which patterns already exist in the workspace.
1. Stay in dialog until it clicks.
Keep the dialog going until the AI has no more follow-up questions and you have no more follow-up questions. Only when the explore step really sits, you move on.
2. Then send both instructions together in one message:
3. After that: /shutdown. Close this terminal, open a new one.
New terminal with cr. Paste the prompt from step 2. Claude writes an implementation plan in detail: file by file, step by step, validation checklist, success criteria.
1. Stay in dialog until it clicks.
You go through the plan with the AI until both sides are satisfied. No gaps, no open questions.
2. Then send both instructions together in one message:
3. After that: /shutdown. Close this terminal, open a new one.
New terminal with cr. Paste the implementation plan. This is where the real magic happens: files get written, code gets built, tests run.
1. Stay in dialog until it clicks.
You guide the build, give feedback, correct. Iteration by iteration, until Claude says: I'm done.
2. Check it yourself.
Once Claude says the implementation is complete, you verify on your own that it really fits. Does it work? Does it look the way you want? If not: name your changes in the chat, again and again, until you're really satisfied.
3. After that: /shutdown. Done.
When you work with Claude's context window, one thing matters most: never max it out. Especially when you build something important, a new project that runs over a longer time.
That's why every time: /shutdown, close the terminal, open a new terminal. That way you always pull the maximum out of every session.
Shape the idea, clarify options. Output: Markdown.
Deepen discovery, clarify constraints. Output: Markdown.
Implementation plan in detail. Output: Markdown.
Run the plan. This is where building happens.
After every step. Clean handoff to the next.
Each step starts in a fresh terminal with cr.