Best-Model-First Agent Workflow
Use the most capable model on max effort, start in plan mode, then auto-accept — counterintuitively cheaper.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 88%
A concrete operating procedure for getting the most out of a coding agent. The core counterintuitive move: always use the most capable model at maximum effort, because a cheaper, less intelligent model burns more tokens through correction and hand-holding. Then anchor every task in plan mode (a single injected instruction to not write code yet) and auto-accept edits once the plan is right.
Origin
Boris Cherny's personal Claude Code usage tips, shared as one valid path among many (he stresses there is no single right way).
Core principles
- 01The most capable model is often cheaper overall because it needs fewer tokens and less correction
- 02Plan mode is literally one injected sentence: 'please don't write any code yet'
- 03A good plan lets the model one-shot the execution, so auto-accept becomes safe
- 04There is no single correct workflow — different builders need different interfaces
- 05Meet yourself where you work — terminal, desktop, mobile, or Slack all run the same agent
How to run it
- 1
Default to the most capable model at max effort
Use the strongest available model with maximum effort enabled, rather than reaching for a cheaper model to save money.
Pro tip A weaker model often takes MORE tokens to finish the same task, so 'cheaper' models can cost more overall.
Watch out Don't assume a less expensive model is actually cheaper — measure total tokens to completion, not per-token price.
- 2
Start in plan mode
Begin roughly 80% of tasks in plan mode so the model proposes an approach before writing code. In the terminal this is shift-tab twice; buttons exist in desktop and web.
Pro tip Plan mode is just one injected sentence — nothing fancy — so use it liberally.
- 3
Iterate the plan, then let it execute
Go back and forth until the plan is right, then hand execution to the model.
- 4
Auto-accept edits
Once the plan is solid, auto-accept edits — with a strong model it will one-shot the implementation nearly every time.
Watch out Auto-accept is only safe after the plan is genuinely good; skipping planning removes the checkpoint that makes it reliable.
- 5
Try different interfaces
Experiment across terminal, desktop app, iOS/Android, and Slack — it's the same agent everywhere. Find the surface that fits you.
Pro tip Boris now splits his coding roughly a third terminal, a third desktop, a third mobile.
In the wild
Boris ships 10-30 pull requests a day with 100% of his code written by Claude Code, and runs about five agents in parallel. He starts ~80% of tasks in plan mode, then auto-accepts, and codes roughly a third in the terminal, a third in the desktop app, and a third on the iOS app.
→ A high-throughput workflow that let him not hand-edit a single line of code since November.
Common mistakes
Reaching for a cheaper model to save money
A less intelligent model needs more correction and more tokens for the same task, so it is often more expensive and slower, not cheaper.
Auto-accepting without planning first
Auto-accept only works because a good plan lets the model one-shot the task; skipping plan mode removes the checkpoint and produces unreliable output.
Is it for you?
Best for
Developers and non-engineers actively using Claude Code or similar coding agents who want higher throughput
Not ideal for
Users on strict per-token budgets who cannot use the top model, or teams needing heavy manual review of every line
From the transcript
“just use the most capable model”
“it's actually cheaper and less token intensive if you use the most capable model”
“use plan mode. I start almost all of my tasks in plan mode, maybe like 80%.”
“I auto-accept edits after that cuz if the plan works good, it's just going to one-shot it.”
From the episode
Head of Claude Code: What happens after coding is solved
Boris Cherny