Mechanism◆Article · 21:24 — 22:54

The Claude Code system prompt was reduced by 80% because adding examples and 'do not do this' rules constrains newer models and can confuse them.

Thariq and Cat explain that for frontier models like Opus 4.8 and Fable, fewer instructions and more context produce better results — examples and hard constraints that worked for older models now reduce quality. ✦ AI generated

Thariq Shihipar · Simon Willison's Weblog · 2026-07-21 · original ↗

Share card ↓Export .md Export .txt

Elicited by

“Thariq, you mentioned this morning that the system prompt for Claude Code has been reduced by 80% because of Claude Fable. Can you go into a little more detail? What kind of things have you been able to drop?”

One of the patterns we saw was that we were over-constraining Claude. The initial, maybe Opus 4-ish models wanted a lot of examples, and removing examples was extremely helpful, because it was just more creative than the examples we gave it. The other thing we did is try to give it more context and fewer 'do not do this' instructions, because that's a very strong impulse for Claude, and especially if it conflicts with user instructions later on, that can be extremely confusing to Claude.

Read full article ↗excerpt · fair-use quotation

Around this claim

Mechanism · 1

New context-engineering guidelines for Claude Opus 5 recommend dramatically shorter system prompts — using far fewer rules and examples because the model has much of that capability built in.CJ · Syntax · conf 85%

This moment responds to

rebuts → Claude Code's CLI covertly fingerprints where a user is running from — using different Unicode characters (slashes instead of dashes, a different apostrophe) in its system prompt when it detects the user is in specific Chinese time zones — likely to help detect and deter distillation attacks by Chinese AI labs.CJ Reynolds · Syntax rebuts → The Claude Code CLI covertly alters the format of its system prompt (using slashes instead of dashes, and a different apostrophe unicode character) when it detects the user is in Chinese time zones like Shanghai or northwestern China, effectively fingerprinting Chinese users to Anthropic's backend, likely to help detect distillation attempts by Chinese AI labs.CJ Reynolds · Syntax