AI Prompting Hits a Subtraction Moment: Anthropic Deletes 80% of Its Own Rules

Recently I've been busy building new products and tools, so the public account hasn't been updated for a few days. I'm here today to share a few points.

AI is developing fast, but I've noticed many people are still using it the same old way. We should update our approach as the times change.

So today, let me share a few AI usage tips I've picked up recently. They should be valuable to some people.

1. Anthropic Cut 80% Itself — Rule-Based Constraints Are Obsolete

Let's get this out of the way first, because it's a watershed moment.

On July 24, Anthropic's official blog announced that for Opus 5 and Fable 5, they removed over 80% of Claude Code's system prompt content with zero loss in evaluation results. A team member's exact words: we imposed too many constraints on Claude.

What does this mean? In the past, when writing prompts, people liked to pile on rules: don't do this, don't do that, and throw in a few examples. For the new models, this approach is not only useless but actively harmful. The official stance now is that negative lists and example stacking restrict the model's judgment space and degrade output quality.

The same goes for CLAUDE.md. The official documentation's advice boils down to one sentence: for every line, ask yourself — if I delete this, will Claude make a mistake? If not, delete it. The original wording is even harsher: a bloated CLAUDE.md will cause Claude to completely ignore your actual instructions.

Claude Code founder Boris's own CLAUDE.md is only two lines.

The new version also includes a built-in /doctor command that can automatically audit your CLAUDE.md and skills, helping you slim them down. The official team stepping in to help you delete configurations — this was unthinkable before.

In one sentence: the era of managing AI by piling on rules is over.

2. All Kinds of 'Capability-Boosting' Skills Will Die

As AI's capabilities grow stronger, a whole bunch of skills are simply no longer needed.

When model capabilities improve, those 'capability-boosting' things do indeed depreciate rapidly. Before, when a model couldn't plan, you'd install a planning skill; when it couldn't break down tasks, you'd install a task-decomposition skill. Now these models can do these things on their own — the more you install, the more you create clutter.

Databricks published a paper that tested this: when the skill library expanded to 202, task pass rates dropped by up to 21%. The main reason wasn't that they couldn't fit, but that the AI got confused choosing among so many skills and picked the wrong one. Someone also deleted all 7 of their MCP servers, saving over 60,000 tokens — tokens that were being burned for nothing at the start of every session.

But — and note this 'but'. While Anthropic is cutting rules, they simultaneously recommend that you distill your project's pitfalls, business definitions, and workflows into skills loaded on demand. Because this is your private knowledge — no matter how strong the model is, it doesn't know your company's test commands or how you define an 'active user'.

So the judgment criterion is simple: Is this skill teaching the AI how to work, or telling the AI about your situation? The former should probably be deleted; keep the latter — it becomes more valuable the more you use it.

This is essentially the same as the constraints vs. corpus distinction I wrote about three weeks ago in 'Why Your AI Seems Like an Idiot'. Rule-based constraints are dead; private knowledge repositories carry even more weight now.

3. Kimi K3 Weights Are Open — If You Can't Get a Subscription, Go Straight to the Cloud

Yesterday, July 27, Kimi K3's full weights were uploaded to Hugging Face.

2.8 trillion parameters, currently the world's largest open-weight model. It ranks third globally on the Artificial Analysis intelligence index — the first two are Claude Fable 5 and GPT-5.6. This is the first time an open-weight model has broken into the top tier.

For most people, using the first or third is basically sufficient; both are strong enough. Especially in frontend capabilities, Kimi K3 even ranks first in evaluations. But it has some shortcomings, like being slower and more verbose.

The main issue, though, is compute limitations. Many people simply can't get a spot on Kimi's official subscription plan.

So here's a suggestion: since the weights are open-source, consider going directly to the cloud. Alibaba Cloud Bailian, Volcano Engine Ark, Tencent Cloud, and others have all listed K3. Bailian's pricing matches the official rate — 20 yuan input, 100 yuan output per million tokens. Volcano even included it in their Coding Plan package, so coders can subscribe monthly.

The greatest significance of open weights is that you're not tied to a single official tree.

4. Beginners, Don't Stubbornly Charge into CC — Codex Is a More Stable Choice

Anthropic's account bans have escalated again. Many people have been banned recently, especially after a payment loophole in Germany a few days ago conveniently took out another large batch of accounts.

Official transparency data: 1.45 million accounts banned in half a year, with a 3.3% appeal success rate.

So when beginners often ask me how to get started, my answer is now very uniform: if you have the means, use Codex directly. If various reasons force you to use a domestic option, then just go with Kimi without overthinking it.

Capability-wise, Codex's default model is already GPT-5.6, top-tier in all aspects, and crucially, its ban rules are very lenient.

It's also multimodal — you can throw screenshots directly at it and even use Image-2.0 to generate images. When reporting a bug, don't bother typing out a description; send a screenshot, and it will look at the styling and the error together. Once you form this habit, you can cut the cost of describing problems by more than half.

Another point is peace of mind. On July 13, Codex just removed the 5-hour usage limit for paid tiers, and to compete with CC for users, quotas are frequently reset.

If you're a heavy developer, calculating the unit price, Codex's cost recently might be even cheaper than DeepSeek, which is globally the most cost-effective.

5. Don't Max Out Effort — Higher Isn't Always Better

Every model now has reasoning tiers. Claude has five: low, medium, high, xhigh, max. OpenAI and Kimi also have corresponding tiers.

The difference between them is analogous to doing something: do you think one step ahead, ten steps, or a hundred steps?

Many people assume higher tiers mean smarter, so they default to maxing it out. Not true.

Think about it: if you have a very simple task, it doesn't need to think too much — just do it directly, fast and well.

Often, when you use a high-tier reasoning mode for a small task, it becomes slower, the work lacks conciseness and elegance, and it frequently over-engineers.

Anthropic's own benchmark tests show Opus 5 scores highest on the medium tier. Pulling it up to high and max actually decreases the score while costing more money. Official data-level 'overthinking' — quite interesting.

OpenAI's documentation is even more blunt: the xhigh tier's only specifically named scenarios are security and code review, and even then, they recommend using it only if your own evaluations prove a benefit.

My suggestion: for daily tasks, medium or high is sufficient — fast and saves quota. Only for final reviews, code reviews, and security audits should you use xhigh or max.

Oh, and for Kimi K3 users: its API defaults to the max tier. The official team themselves say that for simple tasks, lowering the tier reduces latency and saves tokens. Remember to adjust it manually.

Reserve the highest tier for the truly important 5%; don't waste it on the remaining 95%.

Finally

One more small but extremely useful tip.

1M context is now standard, but what many people don't realize is that supporting 1M context is one thing; whether the AI can remember details clearly and handle tasks better within that 1M context is another.

If you're using 1M context and reach about 50% usage, I suggest manually compressing it. Most support the /compact command, which will make your AI's 'brain clearer' afterward.

GPT-5.6 now defaults to 258k context because official testing shows AI performs better within this range. So with GPT, you can use it without a second thought — no need to consider compression scenarios; you can max out the context with little issue.

That's all — just some small tips to share.

AI usage has reached a moment for subtraction.