Source: Simon Willison's Weblog — 2026-07-21
Summary
Simon Willison published an annotated transcript of a fireside chat with Cat Wu (Head of Product) and Thariq Shihipar (Engineer), both from Anthropic's Claude Code team, recorded at the AI Engineer World's Fair. The headline finding runs against conventional prompt-engineering wisdom: as frontier models got smarter, Anthropic found that loading the system prompt with examples and long "don't do this" lists actively worked against Fable 5-class models, and cutting Claude Code's own system prompt by 80% for the newest models made behavior more reliable, not less.
Key Takeaways
- Claude Code's system prompt was cut by roughly 80% specifically for frontier models (Fable 5, Opus 4.8) — older models still run the full-length prompt, meaning Anthropic maintains model-tier-specific prompt lengths rather than one prompt for all.
- The stated reasoning: examples in a system prompt don't just inform smarter models, they constrain them — a model capable of generalizing gets anchored to the literal pattern of the examples instead of reasoning about the actual task in front of it.
- Concrete product metric: Claude Code's Slack integration is already landing 65% of the team's own product-engineering pull requests, offered as evidence the shorter-prompt, less-constrained approach holds up in real internal usage, not just benchmarks.
- The conversation covered tool design, evals, and coding-agent security as related threads, with Anthropic describing its own internal usage of Claude Code as a forcing function for what ships externally.
- The practical implication for anyone writing agent system prompts: the reflex to add more examples and more explicit constraints as a model gets smarter may be backwards — the newest models may need less scaffolding, not more, and over-specifying can cap capability you're already paying for.
Reel Script
Hook (~17s, 39 words) Most teams handle a smarter model by writing a longer, more detailed system prompt. Anthropic did the opposite with Claude Code — they cut the system prompt by 80% for their newest models, and it got more reliable, not less.
Core Concept (~60s, 132 words) A system prompt is the standing instructions an AI agent reads before every single task — the rules, examples, and constraints that shape how it behaves. The default instinct as models get more capable is to add more of this: more examples of correct behavior, more explicit "don't do X" guardrails, more edge cases spelled out. Anthropic's Claude Code team found the opposite happening with their frontier models. Once a model is smart enough to generalize the underlying task, feeding it a pile of literal examples doesn't teach it the pattern — it anchors it to that exact pattern, narrowing behavior instead of guiding it. Fewer explicit constraints let the model apply its own judgment, which for a strong-enough model produces better outcomes than a heavily scripted prompt.
Hands-On (~50s, 112 words) The concrete change: Claude Code maintains different system prompt lengths per model tier — the newest frontier models get roughly a fifth of the token count that older models still receive, with the bulk of the cut coming from removed examples and constraint lists rather than removed capability descriptions. Anthropic backs this with a real usage number, not just a benchmark: Claude Code's own Slack integration is landing 65% of pull requests across their product-engineering team using this leaner prompt. That's the test that matters — not "does it score well on an eval," but "does it actually ship working code inside the team that built it," at a rate high enough that it's now a normal part of their workflow.
Takeaway (~22s, 48 words) If you're hand-tuning a system prompt by piling on examples every time the model underperforms, you may be fighting your own model's capability instead of using it. Try trimming before you try adding — especially on frontier-tier models. Test the leaner version against a real internal workflow, not just a benchmark.