Anthropic Cuts 80% of Claude Code Prompts, Opus 5 Adds 72%

Claude Code just announced a significant reduction in prompts, and immediately after, Opus 5 was caught in a "rebound"!?

And this rebound was actually quite significant: 72%. (laughs)

These seemingly contradictory figures come from Anthropic’s official data and a netizen’s real-world test.

A few days ago, Anthropic announced the removal of over 80% of the system prompt content from models such as Opus 5 and Fable 5.

Right after the result, netizen Chen Cheng (@chenchengpro) happened to capture a packet—

The guy extracted and counted the actual system prompts received by different models, and found that Opus 5 was 72% longer than Opus 4.8??

One dropped by 80%, while the other rose by 72%.

Company A, are these prompts really about cutting down, or are you secretly snacking behind everyone’s back...?

Let’s first discuss exactly what Anthropic deleted this time.

According to the official article, early versions of Claude Code included numerous explicit, even absolute, rules in the system prompt to prevent the model from making errors.

For example, do not write comments or multi-line docstrings unless the user requests them, and do not create plans or analysis documents unless specifically asked.

In other words, wherever the model is likely to fail, humans step in ahead of time with a rule!

Over time, the system prompts behind Claude Code became increasingly like a thick employee handbook.

This approach was certainly effective with earlier models, but as Claude's capabilities improved, issues gradually emerged...

The most direct issue is that commands from different locations can easily conflict with each other.

Faced with these overlapping or even conflicting pieces of information, Claude must expend effort just to determine whose advice to follow—even before starting the task!

So, Anthropic completely overhauled the prompts for Claude Code.

One very straightforward change occurs in the comments and documentation guidelines.

The long-standing rules that once prohibited comments and absolutely forbade multi-line docstrings have now been condensed into a single sentence—

Write code that matches the surrounding code style, maintaining consistency in comment density, naming conventions, and idiomatic usage.

Just take a look at how this project was originally written, then make your own judgment based on the actual situation.

Of course, there are many similar changes.

The result was that Company A deleted over 80% of the system prompts in next-generation models such as Opus 5 and Fable 5...

Notably, according to Anthropic’s programming benchmarks, this significant reduction did not result in any measurable performance decline, so there’s no need to worry.

But I still have to ask—didn’t you cut the prompts by 80%? How is the model’s performance exactly the same? Really??

Along with this major overhaul, Claude Code’s context engineering rules have been almost entirely rewritten.

For developers using Claude Code or building agents based on Claude, Anthropic has also provided several more specific adjustment recommendations.

I’ve also put together a quick overview for everyone—

Simplify CLAUDE.md: retain only the project's special rules; remove information Claude can derive from the code.

Break down long instructions into skills: Save processes such as code review, testing, and deployment separately, and load them as needed.

Remove duplicate instructions: Write each rule only once; do not repeat content already covered in the tool descriptions within the system prompt or CLAUDE.md.

Reduce tool invocation examples: Clearly define parameter meanings, optional settings, and return results, and minimize fixed invocation examples.

Provide executable references directly: Developers can give Claude test cases, existing code, HTML prototypes, or other project files for reference, reducing the need for lengthy textual descriptions.

In summary, Anthropic’s new approach can be summarized in one sentence—

Give Claude fewer detailed employee handbooks and instead clearly design the tools, environment, and boundaries of tasks!

What a massive purge of prompts...

Seeing this, Claude Code’s massive prompt reduction appears to have been quite successful.

But a developer on the internet, fixated on Anthropic's claim of "over 80%," suddenly posed a very straightforward question—

Did you really delete that many? (Good question.jpg)

That’s right—netizen Chen Cheng (@chenchengpro) directly pointed the Claude Code CLI at a local server and extracted all the system prompts actually received by different models.

Then, a rather dramatic set of data emerged:

Opus 4.7: 15225 characters

Opus 4.8: 4467 characters

Opus 5: 7694 characters

Yes, upgrading from Opus 4.7 to Opus 4.8, Claude Code's system prompt was indeed subjected to a major overhaul.

The original prompt of 15,225 characters was reduced all the way down to 4,467 characters, significantly trimmed down.

However, in the latest Opus 5, the plot takes a sudden turn—

The actual system prompt length increased from 4,467 characters to 7,694 characters, which is approximately 72% longer than Opus 4.8!!!

So, the significant reduction in Claude Code's prompts does exist, but this large-scale trimming primarily occurred during the Opus 4.8 phase.

After the release of Opus 5, Anthropic added back a lot of content into the streamlined prompt.

Explore again! Dig deeper!

Scrolling further down Claude Code’s implementation, attentive netizens discovered another key detail—

Each version actually includes two sets of prompts, and a function determines which set to send to the model based on the model ID.

Sonnet, Haiku, Claude 3, and some older Opus models use the more detailed legacy prompts; newer models such as Opus 5 and Fable 5 use significantly streamlined new prompts.

In other words, Claude Code has indeed removed many old rules from the new model.

It wasn't until Opus 5 that the model developed some new behavioral patterns, requiring Anthropic to reintroduce more targeted constraints.

In that case, the official claim of “removing over 80%” and the developer’s finding of “Opus 5 increasing by 72%” can both be true simultaneously.

The first number describes the overall transition of Claude Code from the old prompt system to the new one; the second number compares Opus 5 with its immediate predecessor, Opus 4.8.

Even after recovering part of it, Opus 5's 7,694 characters are still only about half of Opus 4.7's.

Then the question arises:

What did Anthropic add to the additional thousands of characters in Opus 5?

Upon further breakdown, careful netizens discovered that two exclusive Opus 5 features were primarily added:

Delivering work: Specifies how to control the scope of tasks, report progress, and deliver outcomes.

Corrections: Prevent the model from repeatedly explaining and correcting its previous errors.

The two parts combined total approximately 3,755 characters, covering the difference between Opus 5 and Opus 4.8.

As for why Opus 5 restored this content, it’s primarily related to the model becoming more proactive—

According to Anthropic's prompt guidelines, Opus 5 prefers to report progress while executing tasks, typically generates longer responses and documents, and is more willing to invoke sub-agents, expand the scope of tasks, and repeatedly verify results.

These habits can indeed be helpful when tackling complex, long-term tasks.

But if a user only wants to modify a few lines of code, Opus 5 still deploys a squad of sub-agents and performs multiple rounds of checks, causing token consumption and wait times to rise accordingly…

Therefore, the new prompt includes more specific requirements: when to invoke sub-agents, how frequently to report progress, at what point to stop the task, and how to control the final output length.

At this point, the previously seemingly contradictory data becomes understandable.

What Claude Code removed were mostly cumbersome operational guidelines left over for older models, while the 72% increase in Opus 5 is primarily used to constrain the proactive behaviors of the new, more powerful model.

The more a model can work independently, the less humans need to guide it step by step. However, as models become more autonomous, it’s still essential to set clear boundaries in advance about what tasks they can initiate and where they should stop.

Looking at it this way, Claude Code’s prompts have indeed gotten leaner, but Opus 5 has just gained weight elsewhere.

Sure enough, even AI can't escape this pattern: when one old issue fades, a new one precisely pops up elsewhere. (doge)

Reference link:

[1]https://x.com/chenchengpro/status/2081339012632215684

[2] https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models

This article is from the WeChat public account "Quantum Bit," authored by: Focused on Frontier Technologies