Plan-then-execute coding

One-line summary: A two-phase agentic coding workflow — iterate the agent to a good plan first, then switch to auto-accept and let it execute — underwritten by the claim that "once the plan is good, the code is good."

The insight

With sufficiently capable models, planning is the bottleneck, not execution. Time spent steering the agent mid-execution is often wasted; time spent steering it during planning compounds. So the human's job is to verify and iterate on the plan, then get out of the way.

Evidence

• From 2026-04-21-boris-claude-techniques (Boris Cherny, creator of Claude Code):
  • "Most sessions start in plan mode. If my goal is to write a pull request, I will use plan mode, go back and forth with Claude until I like its plan. From there, I switch into auto-accept edit mode and Claude can usually one-shot it."
  • "Once the plan is good, the code is good."
  • "Planning is just the most underused feature in Claude Code."
  • Plan-mode capability improved step-change with claude-opus-4-5: "this is definitely not the case with previous models."
• Boris compares the practice to "spec-driven development" but rejects the rigid framing: "there are some kind of spec … I think it's just like a plan. That's all it has to be. It's just a text file. It doesn't have to be in a particular format."

How Boris operationalizes it

1. Open a tab, ask Claude to make a plan for task A.
2. While it's planning, open a second tab, start task B's plan. Third tab, task C's plan.
3. Circle back to task A. Review and adjust the plan. Repeat until the plan is good.
4. Flip to auto-accept edits. Let Claude execute.
5. Move to the next tab. "Tending to the Claudes."

This directly couples plan-then-execute with parallel-claude-workflow.

Design implications

• Plan mode deserves to be the default entry point for non-trivial tasks. A Claude Code configuration that opens in execute mode wastes the biggest lever.
• Plan artifacts are cheap and reusable. They're plain text, so they can be version-controlled, pasted into issues, and fed to other agents.
• The "good plan" threshold is the new review surface. Plan review replaces line-by-line code review for agent-driven work.

Contradictions / tensions

• None surfaced in source. Boris concedes some people still find plan mode unused; his framing is that the underuse is unwarranted, not that there's a counter-camp.

Open questions

• How tightly is this pattern coupled to current Opus-class model capability? If a weaker model is used, where does plan-then-execute break down?
• Is plan quality a learnable skill, or mostly a property of the task-framer's domain expertise?

Sources

• 2026-04-21-boris-claude-techniques

Related

• boris-cherny
• claude-code
• claude-opus-4-5
• parallel-claude-workflow
• agent-output-verification