Viewpoint
Give high-level tasks with guardrails and exit criteria to modern models.
Give high-level tasks with guardrails and exit criteria to modern models.
For modern models, give high-level tasks with guardrails and exit criteria, not overly specific instructions.
- Interview
- Boris Cherny: Building Claude Code
- Speaker
- Boris Cherny
- Topic
- AI agent usage
- Source timestamp
- 14:47
More from this interview
- Opus 5 runs for weeks or months with Auto Mode and isn't prompt injectable.
- Deleted 80% of system prompt because Opus 5 is more intelligent.
- Ablation shows model is smarter without prompts, but product prompts help usage.
- Defines product overhang and hobbling in AI products.
- Claude Code emerged from Sonnet 3.5's untapped file-writing capability.
- Giving models hard tasks with verification is the key skill.
- Simple prompt to rewrite Electron app in Swift runs over two weeks.
- Dynamic workflows let Claude orchestrate thousands of agents efficiently.
- Routines let Claude maintain its own codebases and automate maintenance.
- Coding is solved for some tasks, not deep systems or UI verification.
- Practical application and business sense should complement computer science theory.