Most developers using AI coding assistants fall into the same trap: they type a request, let the model start editing files immediately, and then spend twice as long fixing the mess. Developer Johanna Larsson has a different approach, and it is worth studying.
Her system breaks every Claude Code session into three strict phases, and the discipline she applies between them is what makes it work.
Phase 1: Context Loading, Not Prompting
Before asking Claude Code to do anything, Larsson feeds it everything relevant - Linear tickets, Notion docs, code files, repo structure. The goal is building what she calls a "shared picture of the world" between her and the model. This is not the same as just pasting in a file path. It means giving the AI enough background to understand the problem the way you understand it.
Phase 2: Investigate, Do Not Touch Anything
This is where most people skip ahead, and where Larsson is most strict. After context loading, she explicitly tells Claude Code: no edits, no actions, not even plan mode yet. The model investigates the problem space and reports back. If it tries to jump to a solution, she stops it and demands it explain what it found first.
The reasoning is simple. Language is ambiguous. When you say "fix the auth flow," you and the AI might picture completely different things. Forcing an investigation phase surfaces those misalignments before any code gets written. It is cheaper to catch a misunderstanding in a conversation than in a broken pull request.
Phase 3: Execute, But Reset When Things Break
Only after alignment does she enable auto-accept mode and let Claude Code run. The critical habit here: when something goes wrong during execution, she does not let the model attempt a quick patch. She pulls it back to investigation mode to understand the root cause before trying again.
She also runs six or more concurrent Claude Code sessions, switching between them while each one works. Context-switching becomes productive rather than wasteful because the AI handles the waiting.
The broader point Larsson makes is that the models will keep improving, but the human skill of setting direction and catching misalignment will not become less important. AI coding tools are force multipliers, not autopilots. The developers who treat them like a junior engineer they need to manage - giving clear briefs, reviewing plans before execution, refusing sloppy shortcuts - will consistently get better output than those who type a prompt and hope for the best.