Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic
Summary
Anthropic scaled Claude coding from 2,500 to 30,000+ daily changes by shifting developers from approval gatekeepers to session managers. Modern best practices now focus on trimming verbose prompts and configuring auto-mode permissions, since newer models like Opus 55 need less steering and can verify their own output.
Key Takeaways
- Remove urgent language and detailed instructions from claude.md files—Opus 55 understands context from code alone. Only include information Claude cannot see, like legacy service details or external dependencies.
- Auto mode now handles permission gating by running a classifier on unsafe commands (deployments, external calls), eliminating manual approvals. Configure it in plain English settings to trust specific repos, buckets, and domains.
- Shift from 1-2 concurrent sessions requiring constant human oversight to dozens running simultaneously for hours. The model and harness now own steering and verification instead of developers.
- Run 'doctor prompt audit' command to automatically optimize your claude.md and skills for the current model version, removing now-unnecessary guidance that causes the model to go off-track.
- Modern models excel at long-running tasks and self-verification—Opus 55 as daily driver, Fable 52 for output verification. This enables sessions to orchestrate other sessions without human intervention.
Related topics
Transcript Excerpt
I had some technical difficulties during my talk this morning. So, I ran back home and I'm just going to record it from my computer instead. So, hi everyone. I'm Lydia. I work on the clock code team at Anthropic. And the models have gotten really good lately. So, we have Opus 55, which is my favorite daily driver. And earlier this week, we also launched Fable 52. And these models are extremely good at longunning tasks and understanding your code base. and especially Fable is also really good at verifying its own output. But as these models get better, the way that we work with them in cloud code also changes. So earlier this year, I don't know if you remember, but sessions kind of used to look like this. So we'd use a model like Opus 45, which honestly for its time was really good. I reall…
More from ai.engineer
- Your SKILL.md Is 70% Stuff the Model Already Knows - Nikolay Petrov, Intel
- Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
- What Your Coding Agent Can't See in the Code - Tom Moor, Linear
- How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp
- How to Get Frontier Coding Results at 70% Lower Cost - Nouha Dziri, Cohere