July 25, 2026
Prompt drama goes minimalist
The new rules of context engineering for Claude 5 generation models
Claude dumped most of its rulebook and the comments instantly turned into a roast
TLDR: Anthropic says its latest Claude models work just as well after removing most of the giant hidden instruction list, suggesting simpler guidance may now be better. Commenters turned that into a roast, joking that AI is one step away from being told "you know what to do" and calling the old style hilariously overengineered.
Anthropic basically walked into the room and announced: we deleted over 80% of Claude Code’s giant instruction sheet, and nothing got worse. In plain English, the company says its newest Claude models are now smart enough that they need fewer babysitting rules and more room to make their own choices. Instead of stuffing the bot with endless "do this, don’t do that" guidance, the new advice is to trust its judgment more often. And yes, the internet had thoughts.
The loudest reaction was a mix of smug vindication and nervous laughter. One commenter, luciana1u, joked that the final form of all this is a system prompt that just says "you know what to do" and somehow works. Another, firasd, absolutely dragged the old style of prompting as "Treaties of Westphalia-length instructions," which is a wildly dramatic way of saying the old setup was absurdly long and fussy. That comment perfectly captured the mood: a lot of people seem delighted that the industry may finally be admitting it has been overcomplicating things.
But there was also a delicious side debate over whether "judgment" should even be a thing we expect from a machine. simonw found it funny that "use your own judgement" is now apparently a real instruction. Then came the ultra-dry drive-by from npstr: "The bitter lesson." Translation: bigger, better systems keep beating clever human-crafted rules. And the funniest jab of all came from mycentstoo, who sarcastically suggested inventing a precise language with strict keywords to control the model... then hit us with "Wait a minute..." In other words: the community’s verdict is that AI prompting may be circling back to the oldest programmer joke of all — if you want perfect instructions, congratulations, you’ve reinvented programming
Key Points
- •Anthropic says it removed over 80% of Claude Code’s system prompt for Claude Opus 5 and Claude Fable 5 without measurable loss on coding evaluations.
- •The article defines context engineering as the broader set of context sources beyond the user prompt, including system prompts, Skills, `CLAUDE.md`, memory, and other sources.
- •Anthropic says Claude Code had been overconstrained by overlapping instructions in system prompts, skills, and `CLAUDE.md` files.
- •The company states that many earlier restrictions were originally added to prevent worst-case behaviors, such as deleting files, but are less necessary with newer models.
- •Anthropic’s updated recommendation is to rely less on rigid rules and more on model judgment, supported by newer tools such as memory, artifacts, and skills.