The question: As models improve, do system prompts get shorter?
The data no longer tells a simple monotonic story. Opus prompts more than doubled by April 2026,
then the official-page prompt bodies settled back into the low 3k range before Opus 5.5 grew to roughly 4.1k words.
The latest increase is concentrated in refusal and safety guidance.
⚠ These are the claude.ai published system prompts, not the full API context.
The Opus 4 through Opus 4.7 rows use the Simon Willison raw mirrors; later rows use
Anthropic's published prompt bodies. Category assignments and historical word totals are approximate.
Anthropic splits the Opus 5.5 prompt across two code blocks: 1,877 words before the open <example> tag and 2,231 after it.
A consistent recount of the full Opus 5 and 5.5 bodies, excluding XML tags, gives 3,225 and 4,108 words.
Claude Code is shown separately because its prompt is
hundreds of token-counted fragments, not one comparable word-counted prompt.
Words per category across Claude snapshots
Opus 5 → 5.5: roughly 3.2k → 4.1k words (+27%)
Per-version breakdown
Claude Code prompt surface
540tracked prompt fragments in Claude Code v2.1.198
581ktokens across prompts, skills, tools, and reference data
145ktokens in agent prompts, tool descriptions, and system reminders
Claude Code is not a single system prompt: Piebald's July 1, 2026 snapshot tracks system reminders,
agent prompts, tool descriptions, skills, and reference data injected by feature and context.
Key Findings
Nuanced
Prompts do not shrink monotonically
Opus grew 1,714 → 3,686 words through 4.7, fell to roughly 3.2k by Opus 5, then rose to roughly 4.1k in Opus 5.5.
True
Behavioral workarounds get retired
The behavior section changed shape: explicit patches like letter-counting and puzzle constraints receded, while agentic guidance and tool-discovery rules appeared elsewhere.
True
Safety dominates the newest prompts
Opus 5.5 puts roughly 2.6k categorized words into refusal, child safety, wellbeing, legal/financial scoping, and Anthropic reminder language. Refusal guidance accounts for most of the increase from Opus 5.
Shifted
Identity changes at Fable
Opus identity text stayed fairly stable, but Fable adds the Claude 5 family, Mythos-class positioning, and dual-use safety framing.
Separate
Claude Code is a prompt system
Piebald tracks 540 Claude Code fragments totaling 581k tokens. The core system/reminder bucket is 46k tokens before tools, agent prompts, skills, or reference data.
Shifted
Refusal guidance expanded
Opus 5.5 adds detailed boundaries for weapons, copyrighted visual works, and ambiguous requests. It also removes Opus 5's Fable safeguard-routing and default-stance blocks.