Skip to content

Claude Code

Claude Code cut over 80% of its system prompt, scores held

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →
Image: claude.dev

Anthropic removed more than 80% of Claude Code’s system prompt for Claude Opus 5 and Claude Fable 5 and found no measurable loss on its coding evaluations, Thariq Shihipar of the company’s technical staff wrote in an October 11 post on the Claude blog. If your team’s CLAUDE.md files and skills were written for older models, the post argues they probably overconstrain Claude in the same way, and Anthropic has built its advice into /doctor, a Claude Code command that rightsizes skills and CLAUDE.md files.

What goes in CLAUDE.md now

Shihipar says CLAUDE.md should stay lightweight: a short description of what the repo is for, with most of the tokens spent on gotchas inside the codebase. His example is a project that keeps all its types in one monolithic file. Anything Claude could learn by looking at the file system or the repo stays out. Instructions that only some tasks need belong in a skill. If you have several unique instructions for verifying work, he suggests putting them in a verification skill and referencing that skill from CLAUDE.md.

Skills get the same treatment. The post calls them lightweight guides that help Claude find information when it needs it, and says they should not be overconstrained “except in highly important areas.” Long skills should be split across many files. They work best when they encode opinions, knowledge or practices particular to you, your team or your product. The myth Shihipar names is the central repository: one file holding every known practice, on the theory that Claude would not find it otherwise. A tree of files that load when needed replaces it.

Two older habits go away. The # hotkey that wrote memories into CLAUDE.md gives way to auto-memory, where Claude saves on its own whatever is relevant to the work and to you. And specs no longer have to be markdown plans. A spec can be an HTML artifact, a detailed test suite, a function in another codebase that Claude should port, or a rubric for what good API design looks like, which verifier agents then check work against. Shihipar prefers references written in code and @-mentioned into the session: an HTML mockup of a design will generally produce better results than a description or a screenshot, he writes.

The system prompt is the one layer where the advice depends on what you run. In Claude Code you will likely never modify it. If you are building your own agent harness, the post says this is where you should spend a lot of time. For that case, Anthropic already shipped a command that strips older-model workarounds out of prompts: /claude-api prompt-audit scans the prompts, skills and tool descriptions in your working directory. In a September test migrating from Opus 4.8 to Opus 5 on a customer support benchmark, it cut cost by 14.6%.

Why the comment rule went

When the team read transcripts of Anthropic’s internal Claude Code use, single requests contained conflicting instructions, such as “leave documentation as appropriate” alongside “DO NOT add comments”, because the system prompt, skills and user requests clashed. Claude can usually work out what the user meant, Shihipar writes, but it has to think more carefully before deciding what to do. The old system prompt said:

In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max.

Older models needed that guardrail, because without it the comments they wrote were wrong in many cases. But the rule was itself wrong for users with their own documentation preferences, and for very complex code that needs longer comment blocks. The new prompt replaces it with one line: “Write code that reads like the surrounding code: match its comment density, naming, and idiom.”

Tool guidance changed along the same lines. Giving examples used to be the number one rule for tool use. With the newest models, examples confine Claude to a certain exploration space, so the post recommends putting the effort into expressive parameters instead. In the Todo tool, listing status as an enumeration of pending, in_progress and completed already hints at how to use it, and one instruction to keep a single item in_progress defines the behavior Anthropic wants. Earlier models were more likely to follow instructions near the end of the context window, so some tool instructions appeared both in the system prompt and in the tool’s own description. Now they live only in the tool description.

Much of the remaining cut came from progressive disclosure, Anthropic’s name for loading the right context at the right time. Code review and verification moved out of the system prompt into skills that Claude Code calls when it needs them. Some tools, the Task tools among them, are deferred: their full definitions stay out of context until the agent looks them up through ToolSearch.

Can you see the cut?

The 80% figure covers Opus 5 and Fable 5, and its support is only Anthropic’s own coding evaluations, on which the company measured no loss. Changes to the prompt itself can be checked from outside. In August 2025, Mario Zechner released cchistory, a tool that shows the differences in system prompts and tool definitions between any two Claude Code versions. Zechner noted that such changes show up as Claude using its tools differently, sometimes less efficiently and sometimes more.

Claude Code also has more than one prompt. Alongside the main system prompt there are sections added depending on environment and config, descriptions of built-in tools like Write, Bash and TodoWrite, separate prompts for built-in agents like Explore and Plan, and utility prompts for jobs like conversation compaction and session titles. Piebald’s GitHub repository extracts all of them from each npm release, more than 500 strings in total. By July 16 its changelog tracked prompt changes across 238 versions going back to v2.0.14, and its authors say it updates within minutes of every Claude Code release.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.