Plan & Context

Always begin new features in Plan mode. Iterate on the approach, argue with Claude, and build good context before executing.

The author never starts in edit mode for new features. Plan mode is for verifying assumptions, challenging Claude's approach, and building rich context. The code generation part is easy once the plan is solid. Spend real effort here — it pays off exponentially.

A clean, focused context always outperforms one bloated with trial-and-error. Start fresh rather than accumulating failed attempts.

When you keep trying "try this, try that," the context gets polluted and the AI can't distinguish good attempts from bad. Start in Plan mode, validate the approach, then execute in a single clean pass. If something goes wrong, consider clearing context and starting fresh.

Save your context/progress to a local file before ending a session, so you can pick up where you left off.

The "second brain" concept — tell Claude to save the current work, decisions, and progress to a local CLAUDE.md or project notes file. This creates a persistent record that survives session boundaries. Essential for multi-day features.

Don't load all context upfront. Pull in project notes, todos, and domain knowledge only when needed for the current task.

Keep todos and project context in external files. Load them on demand with commands like "load my context from local projects." This lets you context-switch between projects without carrying unnecessary baggage. Works well with project management MCPs (Asana, Jira, etc.).