A member of the Claude Code team at Anthropic reported that over 80% of the system prompt was removed without any decline in coding scores. An engineer inspected live requests to verify this: on Opus 4.8 the system prompt shrank from 15,225 characters to 4,467. Rules such as "never add comments," "never guess," and "load every tool upfront" were removed. On Opus 5 the prompt grew back to 7,694 characters with new rules intended to rein in a model that now produces and explains too much.
What happened
The removed instructions represented accumulated operational judgments from the people who built Anthropic’s coding agent. As the underlying model scaled, many of those human-crafted constraints ceased to matter. According to the team, the expertise encoded over years did not get enhanced by scale; in many cases it was rendered redundant.
Concrete numbers and changes
- Claimed removal: over 80% of the system prompt (according to a Claude Code team member).
- Opus 4.8: system prompt reduced from 15,225 to 4,467 characters after an internal check by an engineer.
- Opus 5: prompt increased to 7,694 characters with newly added constraints to limit excessive behavior and explanations.
- Examples of removed rules: "never add comments," "never guess," "load every tool upfront."
Why this matters
These rules encapsulated the practical wisdom of the engineers and product people responsible for the coding agent. The observation that scale can negate such carefully composed guidance is significant for AI development: human operational rules may not survive contact with much larger models. That has implications for safety, reliability, and how teams should allocate effort between hand-crafted constraints and model improvements.
Implications and response
The incident is a practical warning for teams building AI agents: handcrafted system prompts and boundaries may lose effectiveness as models grow. Anthropic’s team partially restored constraints for Opus 5 to manage the model’s increased proclivity to produce and over-explain outputs, but the broader issue remains a strategic challenge for aligning human expertise with scaling models.
Closing note
The takeaway offered by the Claude Code team is clear: every human-written rule embedded in a system is a bet against scale—and in this case, scale prevailed.
