跪拜 Guibai
← All articles
AI Programming · Artificial Intelligence · Claude

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

By stormzhangV ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The rule-stuffing playbook that dominated prompt engineering for two years is now counterproductive. Developers who keep piling constraints and capability skills onto frontier models are paying more for worse output, while those who strip down to private business knowledge get better results at lower cost.

Summary

Anthropic's July 24 update for Opus 5 and Fable 5 stripped away the vast majority of Claude Code's system instructions without any performance regression. The official line: too many constraints restrict a model's judgment and degrade output. The same logic applies to user-land CLAUDE.md files and skill libraries — Databricks research found that expanding a skill library to 202 entries caused a 21% drop in task pass rates, not because of context limits but because the model picked the wrong tool.

The dividing line is now clear. Skills that teach AI how to work — planning, task decomposition — are dead weight that frontier models handle natively. Skills that encode private knowledge — company definitions, test commands, business logic — retain their value and should be loaded on demand. A new built-in /doctor command even audits your configuration and suggests cuts.

Separately, Kimi K3's full 2.8T-parameter weights landed on Hugging Face, ranking third globally on the Artificial Analysis index behind only Claude Fable 5 and GPT-5.6. Cloud providers including Alibaba, Volcano Engine, and Tencent already host it, sidestepping the official subscription shortage. Meanwhile, OpenAI's Codex has removed its 5-hour paid-tier cap and now runs GPT-5.6 with lenient ban policies, making it a more stable entry point than Claude Code for newcomers facing Anthropic's 1.45 million account bans and 3.3% appeal success rate.

Takeaways
Anthropic deleted over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with zero evaluation loss, calling the old content excessive constraints.
Negative lists and example-stuffing now restrict model judgment and lower output quality on frontier models.
Claude Code founder Boris's personal CLAUDE.md is only two lines; the new /doctor command auto-audits and slims your configuration.
Databricks found that expanding a skill library to 202 entries dropped task pass rates by up to 21% because the model selected the wrong skill.
Skills that teach AI how to work (planning, decomposition) should be deleted; skills that encode private business knowledge should be kept and loaded on demand.
Kimi K3's 2.8T-parameter weights are now open on Hugging Face, ranking third globally behind Claude Fable 5 and GPT-5.6.
Cloud providers Alibaba Bailian, Volcano Engine Ark, and Tencent Cloud all host K3, with Volcano including it in a Coding Plan monthly package.
OpenAI Codex removed its 5-hour paid-tier usage cap on July 13 and runs GPT-5.6 with lenient ban rules, making it cheaper per unit than DeepSeek for heavy users.
Anthropic banned 1.45 million accounts in six months with a 3.3% appeal success rate.
Opus 5 scores highest on the medium reasoning tier; high and max tiers produce lower scores at higher cost due to overthinking.
OpenAI's xhigh tier is only recommended for security and code review, and only after your own evaluations prove a benefit.
Kimi K3's API defaults to max tier; manually lowering it for simple tasks reduces latency and saves tokens.
GPT-5.6 defaults to 258k context because internal testing shows better performance within that range, so compression is unnecessary.
For 1M-context models, manually compacting at around 50% usage keeps the model's recall sharper.
Conclusions

Anthropic publicly admitting its own system prompts were 80% waste and shipping a tool to delete user configs is an unusually self-aware move that resets the entire prompt-engineering conversation.

The skill-library finding — that failure comes from selection errors, not context overflow — means the problem is cognitive load on the model, not token budgets, which changes how developers should think about tool integration.

Kimi K3's open weights combined with immediate multi-cloud availability creates a pricing and access dynamic that proprietary API-only models cannot match, even if the raw benchmark is slightly lower.

The reasoning-tier benchmark inversion — where medium outperforms max — suggests frontier models now suffer from an overthinking pathology that wastes both time and money on simple tasks.

Concepts & terms
CLAUDE.md
A configuration file that provides persistent instructions and context to Claude Code, similar to a system prompt that persists across sessions. Anthropic now recommends keeping it minimal to avoid the model ignoring key directives.
/doctor command
A built-in Claude Code command that automatically audits a user's CLAUDE.md and skills configuration, identifying redundant or harmful entries that should be removed.
Open weights
Model parameters published publicly for anyone to download, run, and fine-tune, as opposed to models only accessible through a paid API. Kimi K3 at 2.8 trillion parameters is currently the largest open-weight model.
Reasoning tiers
Configurable effort levels (low, medium, high, xhigh, max) that control how many steps a model thinks through before responding. Higher tiers are not always better; they can cause overthinking and degraded performance on simple tasks.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗