When AI Writes 65% of the Code: Counterintuitive Prompt Engineering Truths from Inside Anthropic

Barely a year ago, Claude Code was just a bullet point in the launch of Claude 3.7 Sonnet. Today it’s a core coding force inside Anthropic. In a recent fireside chat at the AI Engineer World’s Fair, Cat Wu and Thariq Shihipar from the Claude Code team sat down with Simon Willison and revealed a host of counterintuitive practices—from 65% of PRs being auto‑submitted by AI to slashing system prompts by 80%, and the discovery that negative instructions can actually harm output quality. These insights suggest many of us are still approaching new AI with old mental models.

Prompt engineering’s paradigm shift: less is more Most developers still craft verbose system prompts packed with examples and “do not do X” constraints. But Anthropic found that for models like Fable or even Opus 4.8, this is no longer best practice. They recently reduced the Claude Code system prompt by 80% because the model can infer intent on its own; extra instructions just muddy the waters. Even more surprising: listing “don’ts” often degrades results—the model seems to fixate on what not to do and becomes more error‑prone. It’s like teaching a smart new colleague: pointing out key principles works better than handing over a thick manual.

From supervision to delegation: how coding agents reshape the developer role Cat recalled that when Claude Code first launched, developers would monitor every action and frequently reject suggestions. Now, with models like Fable, many features can be one‑shot, freeing people to think about higher‑level product experience. Thariq noted that after AI takes over grunt coding, developers can feel a void (he calls it the “Deep Blue effect,” referencing the ennui chess players felt after losing to Deep Blue in 1997). The antidote, he says, is to set more ambitious goals: when AI handles the details, humans should focus on bigger problems.

Layered trust: where does automation stop? Despite the automation, the Claude Code team still manually reviews critical changes. But the outer layers of the product are increasingly subject to AI‑driven code review, creating a layered trust model. Auto Mode, which lets the agent operate without human intervention for stretches, is the enabling technology behind Claude Tag’s 65% PR submission rate. The team emphasizes that features are first used internally and only released externally if they demonstrate strong retention—a rigorous internal vetting process.

Culture: public collaboration accelerates iteration Inside Anthropic, “dogfooding” is called “ant fooding,” but more striking is their culture of working in public. Claude Tag is deeply integrated into open Slack channels, where everyone can see how AI is used and how it fails. This transparency means errors are caught quickly and best practices spread organically. Anthropic’s success stems not just from model capability, but from a culture that lets AI grow under public scrutiny.

For the rest of us, the takeaways are clear: stop clinging to old prompt engineering habits—try trimming prompts and trusting the model. Let AI handle the outer loops of code, but keep human oversight on the core. And build a team culture where AI usage is visible to all, so everyone learns faster. When AI writes 65% of your code, the key isn’t to panic but to reimagine your role.