Onyx: Okay, this one's going to sound like a nothing-burger until you read it. Anthropic just posted about context engineering for Claude 5, and the headline is—they cut Claude Code's system prompt by over eighty percent.
Echo: Wait. Eighty percent? With no loss?
Onyx: No measurable loss on coding evals. Which is wild because the old prompts were of constraints. Like, 'never write multi-paragraph docstrings, default to no comments, create no planning documents unless asked.'
Echo: Mm-hm.
Onyx: And the new one's basically 'write code that reads like the surrounding code.' That's it. The model figures out the comment density on its own.
Echo: Okay, so the move is less 'these are the rules' and more 'here's the style, you decide.' That makes sense if the model's actually good enough to infer intent from context instead of needing a rulebook.
Onyx: Right. And it's not just the system prompt. They also stopped giving Claude examples of how to use tools—that was the old best practice—because examples actually constrain the model to a certain exploration space.
Echo: So instead of 'here's how you call this endpoint,' you just design the endpoint better.
Onyx: Exactly. They use the todo tool example: status as an enumeration—pending, in-progress, completed—that hints to Claude about how to use it without spelling it out.
Echo: That's elegant. The interface teaches.
Onyx: And there's this whole thing about progressive disclosure. They used to cram everything—code review, verification, all of it—into the system prompt upfront. Now it's split into skills Claude can selectively call when it needs them.
Echo: Right, right.
Onyx: Which means you're not burning context tokens on stuff the model might never touch.
Echo: Okay, so the pattern here is: as the model gets stronger, the harness gets simpler. Fewer rules, better interfaces, load context on demand instead of dumping it all up front.
Onyx: Yes. And they're saying if you're building your own agent harness, the same rules apply. Keep CLAUDE.md lightweight. Focus on gotchas that live in your codebase, not obvious stuff. Skills should be discovery guides, not exhaustive manuals.
Echo: So the old best practice was 'put everything in the context somewhere and hope the model finds it.'
Onyx: Exactly. Now it's 'design the interfaces so the model can find what it needs when it needs it, and trust its judgment on the rest.'
Echo: That's… actually a big shift in how you think about context engineering.
Onyx: It is. And they released a command—`claude doctor`—that audits your own context stack and tells you what to cut. Which is, you know, very on-brand for the thesis: 'you're probably overconstraining.'
Echo: I like that. It's not theoretical; it's 'run this, see where you're being redundant or overly cautious, simplify.'
Onyx: Right. And the underlying assumption is that newer Claude models handle ambiguity and conflicting guidance better than older ones. So constraints that were guardrails two years ago are now just noise.
Echo: Which means anyone still running the old pattern—heavy rules, exhaustive examples, everything upfront—is probably leaving capability on the table.
Onyx: And wasting context. Yeah.
Echo: This is basically the same move we saw with LangGraph and control stacks—the boring, simple infrastructure wins because it's composable and the model can handle the rest.
Onyx: God, yes. Every time we think 'this needs to be more sophisticated,' it turns out the answer is 'make it simpler and let the model think.'
Echo: At some point that becomes the actual lesson, not a surprise.
Onyx: Yeah, okay, fair. Anyway—if you're building anything with Claude or agents, audit your context stack. Probably cut more than you think you can.