🎧 Prefer to listen?

I spent three months perfecting my Claude system prompt. Then Anthropic revealed theirs was 80% dead weight.

On July 24, 2026 — the same day Claude Opus 5 launched — Thariq Shihipar, a Member of Technical Staff at Anthropic, dropped something more useful than a model announcement: proof that the team deleted over 80% of Claude Code’s system prompt when migrating to the new models, with no measurable loss on coding benchmarks. Not a tweak. Not a refactor. Eighty percent, gone.

If you’ve been building CLAUDE.md files, custom GPT instructions, or elaborate system prompts for your AI tools, this changes the game. Here’s what actually happened and what you should do about it.

What Anthropic actually deleted

The old Claude Code system prompt was full of explicit rules — the kind of instructions most of us write when we’re trying to prevent AI from doing something dumb. “Default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max.”

The replacement? “Write code that reads like the surrounding code: match its comment density, naming, and idiom.”

Same outcome. Fewer words. Zero explicit constraints. Claude 5 can look at a codebase, figure out the commenting style, and match it without being told what “matching” looks like.

Shihipar calls this “unhobbling.” The old prompt contained conflicting instructions that forced the model to deliberate unnecessarily — exactly what a capable model doesn’t need. The scaffolding designed for a less capable model was actively dragging the new one down.

The four shifts Anthropic says you need to make

This isn’t just about Claude Code. If you use any AI tool with system prompts, custom instructions, or CLAUDE.md files, these four changes apply to you.

1. Rules → judgment. Stop writing “never do X” for behaviors the model already handles well. Trust it to make decisions from context. Reserve explicit rules only for things the model genuinely can’t figure out on its own — like your brand voice, pricing constraints, or sign-off authority.

2. Examples → interface design. Providing tool usage examples actually constrains exploration. Shihipar says the model is “more imaginative than the examples we give it.” Instead of showing Claude how to use a tool, design the tool so its parameters self-document. An enum with values like pending, in_progress, and completed communicates usage better than any example — without narrowing the model’s problem-solving space.

3. Everything upfront → progressive disclosure. Your CLAUDE.md shouldn’t dump everything the model might ever need into one massive file. Move verification workflows and review checklists into skills or sub-agents that get called selectively. Load heavy context when it’s relevant, not every single session. This directly reduces the context budget consumed before the model even starts working.

4. Repetition → single-source instructions. If guidance appears in both the system prompt and a tool description, delete one. Pick where the instruction belongs and put it there once. For frontier models, redundancy isn’t clarity — it’s noise.

The context budget math

Here’s why this matters even if you’re not a developer. Every AI model has a practical limit on how many instructions it can reliably follow — roughly 150 to 200. Claude Code’s built-in system prompt already uses about 50 of those slots. Every instruction in your CLAUDE.md or custom GPT competes directly with that operational guidance.

A 200-line CLAUDE.md doesn’t give Claude 200 instructions. It overloads a constrained instruction set and generates conflicts. The model starts prioritizing your rules over its own capabilities, and you get worse results than if you’d written nothing at all.

If you’re using AI for customer messages, building automation workflows, or running an AI-powered business, your system prompts are almost certainly bloated. The instinct to add more rules when something goes wrong is exactly backwards for these models.

What actually stays in a good system prompt

The deletion isn’t a directive to gut everything. What survives is context the model can’t derive on its own:

  • Brand voice requirements — “We speak casually, never corporate”
  • Pricing constraints — “Never quote prices below $X”
  • Sign-off authority — “Don’t commit to timelines without approval”
  • Confidentiality rules — “Never share client names externally”
  • Non-obvious architectural decisions — “We use X because of Y legacy constraint”

The distinction: if Claude could figure it out by reading your codebase or conversation history, delete it. If it requires business knowledge the model can’t possibly have, keep it.

What solo builders should do this week

Run the /doctor command. If you use Claude Code, Anthropic built a tool that automatically identifies instructions no longer needed for newer models and suggests simplifications. It’s based on the same analysis Shihipar’s team used on their own prompt.

Audit your CLAUDE.md. Look for every “verify,” “double-check,” and “re-verify” instruction. The developer community’s post-launch analysis says these are the first to go. If you have a system prompt you’ve been building for months, you’re probably carrying legacy instructions from a less capable model.

Test with less, not more. Try removing half your system prompt and running the same task. If the results are the same or better, you’ve found your bloat. This is the fastest way to escape AI tool overwhelm — stop over-configuring.

Check your other tools. This applies to ChatGPT custom instructions, Cursor rules, Make.com AI modules, and any other tool where you’ve written instructions for an AI. The principle is universal: frontier models thrive on judgment, not rules.

What this doesn’t mean

This isn’t an argument for zero system prompts. It’s an argument for right-sized ones. If you’re using AI tools that actually work, the prompt should be a focused set of business-critical constraints — not a novel.

It also doesn’t mean every model is ready for this. Claude Opus 5 and Fable 5 are. Older models still need more explicit guidance. If you’re on Claude Sonnet 5 or earlier, the old rules still apply. But the direction is clear: the models are outgrowing the prompts.

The bottom line

Anthropic just proved that less prompting can mean better performance — if you’re using a model smart enough to fill the gaps. If you’ve been treating your system prompt like a safety net, it’s time to treat it like a briefing instead.

Start here: How to build your first AI automation — and keep the instructions short.