Claude Code System Prompt Optimization: Lessons from Anthropic

Anthropic reduced Claude Code's system prompt by 80% for Fable 5 and Opus 4.8, finding that removing few-shot examples and hard constraints improved performance. Key lessons: deleting examples is effective because models are more creative without them; avoid 'don't do X' bans as they create conflicts with user instructions; ensure every instruction is 100% correct with no exceptions; rely on the model's judgment for high-capability models. The approach uses progressive disclosure via Skills, automatic memory, and rich references (code, HTML, tests) instead of upfront rules. Claude Tag, a Slack-integrated Claude, now lands 65% of Anthropic's product engineering PRs. Code review is automated for non-core areas, with trust built over six months by adding failure cases to evaluation suites. The team emphasizes that this prompt philosophy is model-dependent: smaller models still need detailed prompts. The shift reflects a broader trend from rigid rules to flexible, context-driven guidance.
Key insights
Removing Few-Shot Examples and Hard Constraints Improves Frontier Model Performance
Anthropic found that deleting examples from system prompts was 'extremely effective' because models like Fable 5 are more creative than the examples given. Similarly, reducing 'don't do X' bans prevents conflicts with user instructions, as such bans create strong impulses that clash with later requests. The result was an 80% prompt reduction with no loss in coding benchmarks.
The '100% Correct' Rule Eliminates Ambiguous Instructions
The team scrutinized every line of the system prompt, asking if any exception existed. Many instructions were only 90% correct—e.g., 'always verify frontend changes' fails for trivial text edits. The new standard is to soften wording until the instruction is universally true, or to let the model use its own judgment, because the prompt is fed to the model 100% of the time.
Modular Context via Skills and Progressive Disclosure Prevents Overload
Instead of loading all context upfront, independent information (code review rules, verification steps) is placed in separate Skills files that the model can search on demand. This avoids overwhelming the model and allows context to be fetched only when relevant, aligning with the principle of 'progressive disclosure.'
Claude Tag Automates 65% of PRs, Demonstrating Production Impact
Anthropic's internal Slack bot, Claude Tag, handles routine coding tasks, monitors repositories, and submits PRs. It now lands 65% of the product engineering team's PRs. Complex tasks still use interactive Claude Code, but routine work is increasingly automated, with code review for non-core areas fully handed to AI after six months of trust-building.
Prompt Philosophy Is Model-Dependent; Smaller Models Still Need Detail
The lean-prompt approach works only for frontier models with strong judgment. Smaller models like Haiku still require detailed, prescriptive prompts. This is why Anthropic maintains separate system prompts per model tier, and the team notes that frontier intelligence is 'almost always worth it' even if it sometimes uses fewer tokens on hard problems.
Context Engineering Shifts from Manual Rules to Automatic Memory and Rich References
Anthropic moved from manually written claude.md memories to automatic model-generated memory, and from simple text rules to rich references like HTML templates, test files, and scoring rubrics. This allows the model to infer patterns from actual code and artifacts rather than from abstract instructions.