Claude Code's System Prompt Cut by 80%: Key Insights from AI Engineer World's Fair Fireside Chat

Claude Code product lead Cat Wu and engineer Thariq Shihipar revealed in a fireside chat at the AI Engineer World's Fair that Claude Code's system prompt has been reduced by 80%, and this change applies to Fable 5, Opus 4.8, and future frontier models. This counterintuitive adjustment stems from a fundamental rethinking of prompt engineering methods.

Thariq stated bluntly that early models (Opus 4 era) needed many examples, but now 'deleting examples is extremely effective because the model is more creative than the examples we give it.' Host Simon Willison admitted on the spot that his primary advice to users had been 'give it examples,' and if that no longer holds, his mental model of prompting is 'a bit broken.'

The team also found that using fewer 'don't do X' prohibitions and providing more context works better. Prohibitions are 'very strong impulses' for Claude; once hard constraints in the system prompt conflict with subsequent user instructions, the model becomes confused. Therefore, the direction is: fewer hard constraints, more background information, and fewer total instructions.

Cat Wu provided an actionable methodology: examine every sentence in the prompt and ask yourself, 'Is there an exception scenario for this sentence?' They found that many instructions were only 90% correct—for example, 'must verify after frontend changes,' but if it's just changing a copy string and the user says 'just make a quick change,' forced verification is wrong.

Her criterion is: imagine how a well-intentioned person might misinterpret the prompt, then soften the wording until it is 100% accurate—because this text is fed to the model 100% of the time. Instead of specifying 'when to test,' directly let the model 'use its own judgment to decide.'

Simon pointed out a key prerequisite: this capability is only available with Opus/Fable-level models; models from a year ago did not have such judgment. Therefore, small models (like Haiku) still need detailed prompts—this is why each model has its own system prompt. Thariq also added a counterintuitive observation: large models sometimes use fewer tokens than small models on difficult problems, and 'frontier intelligence is almost always worth it.'

Both confirmed that current good models are very good at writing prompts. A subagent is essentially one Claude writing a prompt for another Claude; the prompt for Claude Code's Workflow tool was itself written by Claude. Simon later put this into practice: he had Fable 'use its own judgment to choose a low-end model to run a subagent for coding tasks,' which worked well and significantly saved Fable quota.

This philosophy extends to tool design: the trend is to subtract—grep/glob have been removed in favor of native bash, keeping tools few and non-overlapping in functionality. Thariq described 'models are more like biology than physics, and tool design is an art.'

The chat also introduced Claude Tag (Slack-collaborative version of Claude): default multi-user collaboration, ability to actively monitor long-term (e.g., automatically fix bugs and submit PRs), and channel-level team memory. The most striking data point: Anthropic's product engineering team has landed 65% of its PRs through it. In terms of division of labor, complex tasks use Claude Code for interactive iteration, while routine tasks are handed to Claude Tag.

Regarding code review automation, core areas are still manually reviewed by code owners, while peripheral changes are fully automated. Trust was built over six months—each incident added the offending PR to the evaluation set to ensure the review system never regresses.

Evaluation is seen as the foundation: new models must run a full evaluation set to confirm they are 'strictly better than the previous generation' before being swapped in; in addition to capability evaluations, behavioral evaluations are being built (e.g., users dislike the model saying 'it's time to sleep'). The bottleneck is not evaluation tools but the skill to build high-quality evaluations.

Auto Mode is used internally by everyone. The mechanism is that each tool call is dynamically judged by a Sonnet classifier combined with conversation context, and it is linked with a sandbox. It claims that the risk of prompt injection and data leakage is 'far lower than manual review'—evaluation data is pending release, and Simon is cautiously optimistic. There is also a credential injection mode: credentials are dynamically injected at request time, making them 'available but invisible' to the agent.

Engineering mindset has flipped: 'rewriting is good' now holds—the codebase is the only spec, and with a good test suite, you can confidently rewrite; the engineer's value center has shifted from execution ability to product taste and business judgment (the time from idea to implementation has shrunk to one week). Facing the feeling of being replaced by models, the answer is to counter with greater ambition—don't just turn old work into a prompt; do things you didn't dare to do before. Cultural motto: 'We don't negotiate with ourselves.' Just do it, and let the trade-offs reveal themselves.