Type: technique
Status: draft
Domain: horizontal
Tags: knowledge, ai-pm
Last updated: 2026-02-19

Plan Mode as Claude Code Default

Summary

Start 80% of Claude Code tasks in plan mode. Plan mode is a single-sentence injection: “don’t write code yet.” That’s the whole mechanism — one sentence that tells the model to plan before acting. The agent works through the problem, produces a plan, and waits for human review before touching any files.

The payoff: once the plan looks right, Boris Cherny auto-accepts edits and Claude Code one-shots the implementation nearly every time with Opus 4.6. The brief upfront pause — reviewing a plan instead of reviewing code — produces dramatically better outcomes than diving straight into implementation. You catch misunderstandings at the cheapest possible moment.

How to Apply

  1. Activate plan mode before describing the task. In Claude Code, this can be done with a slash command or by prefacing your prompt with an instruction not to write code yet.
  2. Review the plan before accepting. The plan is where you catch scope misunderstandings, missing context, or wrong approach — not after 200 lines of code have been written.
  3. Correct the plan, not the code. If the plan is wrong, fix it in prose before the agent touches files. Prose corrections are cheaper than code corrections.
  4. Auto-accept after a good plan. Once the plan accurately reflects what you want, let the agent run without interruption. With Opus 4.6, one-shot implementation is common from a good plan.
  5. Use it for 80% of tasks. The remaining 20% are genuinely simple — clear enough that planning overhead doesn’t pay off. Develop a feel for which tasks need it.

The underlying principle: align on the approach before generating implementation. Plan mode simply makes this structured and enforced. It’s the equivalent of “no coding before design” applied to agent workflows.

Sources

From: 2026-02-19 Head of Claude Code: What Happens After Coding Is Solved

Key quote: “Boris begins 80% of his tasks there. All it does is inject one sentence telling the model not to write code yet. Once the plan looks right, he auto-accepts edits and it one-shots the implementation almost every time with Opus 4.6.” Attribution: Boris Cherny What this source adds: First-person workflow from the creator of Claude Code. Cherny ships 10–30 pull requests daily using this technique — the volume validates the method. The “one sentence” framing demystifies what plan mode is and why it works. Links: Original | Archive


This site uses Just the Docs, a documentation theme for Jekyll.