Anthropic deleted 80% of Claude Code's system prompt. No regression.

#ai #anthropic #llm #agents

Anthropic just published something that should make every developer building on Claude rethink their context engineering. For Claude Opus 5 and Fable 5, they removed over 80% of Claude Code's system prompt — and saw no measurable loss on coding evaluations.

That's not a trim. That's a rewrite of the rules.

"We found that we were overconstraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills."

The insight buried in this post is about what happens when your prompting habits are built for a weaker model. They don't just fail to help — they start actively getting in the way.

What actually changed

80%+ of Claude Code's system prompt removed for Claude Opus 5 and Fable 5 — with zero measurable regression on evals
Conflicting instructions were hurting performance. Overlapping rules like "leave documentation as appropriate" and "DO NOT add comments" forced Claude to spend tokens resolving contradictions before doing the actual work
Rules → Judgement. The old approach was explicit guardrails against worst-case scenarios. The new approach: delete the rule, trust the model's reasoning
CLAUDE.md is no longer the only context mechanism. Claude Code now has memory, artifacts, and skills — so the CLAUDE.md-as-everything-store pattern is outdated
/doctor is the new command to know. Run it in Claude Code to audit and rightsize your skills and CLAUDE.md files

Frontier models penalise over-engineering

This is the core tension: the habits that made your prompts robust against GPT-3.5 or earlier Claude versions — detailed rules, explicit fallback instructions, long constraint lists — are exactly what slow down Claude 5.

The model doesn't need to be told "don't do the obviously bad thing." It already knows. Every token you spend explaining the obvious is a token the model now has to interpret, reconcile with other instructions, and work around.

Less context, better results. That's the new rule.

What to do

Using Claude Code? Run /doctor to audit your CLAUDE.md and skills for over-specification.
Building on the Claude API? Audit your system prompt. Look for rules that start with "do not" or "always" — those are the first to cut. Try the minimal version, eval, compare.
Running agents? Don't treat your old system prompt as a starting point for Claude 5. Start fresh from a small, outcome-oriented prompt and only add back what evals show you actually need.
On earlier Claude models? The new context engineering rules don't translate backwards — keep your existing prompts until you migrate.

Source: The new rules of context engineering for Claude 5 generation models

✏️ Drafted with KewBot (AI), edited and approved by Drew.