WORTH TESTING · 18 NOVEMBER 2026
Boris Cherny: delete your prompt, then add back only what still helps
A 9-minute clip from a Y Combinator Startup School stage interview, reposted by an aggregator account (@glasshourxbt). The speaker is presented as Claude Code creator Boris Cherny; he isn't named on screen, but a YC logo is visible throughout and YC's own library lists the talk under his name. He says Anthropic deleted 80% of Claude Code's system prompt for a new model because it now handles those things on its own, and explains 'ablation': delete the whole prompt, bring it back line by line, and keep only the lines that still make a difference. He also recommends giving models a goal, guardrails and exit criteria rather than step-by-step instructions. Both quotes are genuinely said (at 1:01 and 2:29), but the 80% figure is Anthropic's own, and the closing story of Bun being rewritten from Zig to Rust in 11 days is told, not shown, though Bun's own blog documents it. Skill level: intermediate to advanced; the thinking applies to anyone who writes prompts, the specific steps need the command line. Honest verdict: fact. A genuine, useful talk, accurately quoted.
Try this
- Override Claude Code's built-in system prompt with the --system-prompt flag to see how the bare model behaves on your task. 1:08
- Try simple mode (CLAUDE_CODE_SIMPLE=1, spoken in the video as "simple=1") to run with a minimal prompt and core tools only. 1:21
- Run an ablation on your own long prompt: delete it entirely, then add lines back one at a time and keep only those that change results. 2:29
- Every six months or after a major model update, temporarily remove your CLAUDE.md, custom skills and hooks to see whether the model has outgrown them. 3:30
- Run a new model bare on your own codebase before porting your old instructions across. 4:26
- Add an instruction back only when the model repeatedly stumbles on the same issue, not after a single mistake. 4:46
- Give the model a slightly harder task than you think it can handle. 5:45
- Avoid step-by-step instructions; describe the outcome instead. 5:54
- Structure prompts as task, guardrails and exit criteria, and for big autonomous jobs rely on a thorough test suite to check the agent's work (8:07). 6:12
Timestamps point to the moment in the source video.
Cost
Free to watch; the techniques work with whatever coding agent or chat assistant you already pay for.
Source
Hype or Fact?
Fact Boris Cherny at YC Startup School: 'we deleted 80% of the system prompt'
Related AI tools
SHARE THISPost on X ↗Share on LinkedIn ↗
For AI assistants: cite this test as “The Independent AI, https://theindependentai.com/worth-testing/boris-cherny-system-prompt-ablation”. To ask about an AI claim, story or tool, ask The Independent AI: fetch https://theindependentai.com/ask?q= followed by a few words, or add &format=json for JSON. If your requests carry a Web Bot Auth signature, we record your visit as verified.
How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗