Why Claude Code Deleted 80% of Its System Prompt — and What Builders Should Delete Next
Site Owner
Published on 2026-07-27
Anthropic's Claude Code team revealed they cut more than 80% of the system prompt when targeting Opus 5 and Fable 5 — coding evals did not regress. Here is the six-rule Then/Now shift, and the deletion checklist for your own CLAUDE.md.
Why Claude Code Deleted 80% of Its System Prompt — and What Builders Should Delete Next
Anthropic's Claude Code team just published a counter-intuitive production result: when pointed at Claude Opus 5 and Claude Fable 5, they cut more than 80% of the system prompt — and saw no measurable loss on coding evaluations. The lesson is not "use less prompt." The lesson is that the right move when models get smarter is to delete, not to add.
On July 25, 2026, Thariq — a Claude Code team member at Anthropic — posted a long thread on X titled "The new rules of context engineering for Claude 5 models" (status 2080710971228918066). Inside, the team revealed what had changed inside their own harness as the underlying model moved from Claude 4 to Claude 5 (Thariq, https://x.com/trq212/status/2080710971228918066).
The headline number stops most readers cold:
"We deleted more than 80% of Claude Code's system prompt for Opus 5 and Fable 5, and there was no measurable loss on our coding evals." (Thariq, https://x.com/trq212/status/2080710971228918066)
The number is provocative. Most teams treat the system prompt as scaffolding — every bug found adds a rule, every rule stays. Anthropic went the other way. The reason matters more than the percentage.
A model that can already do the right thing does not need a memo telling it how.
Why the system prompt had grown in the first place
When Claude Code launched, the underlying model was good enough to be dangerous and not good enough to be trusted. It would call the wrong tool, double-search, drop a step. So the team added guardrails. Each guardrail fixed one observed failure. Over time those patches turned into luggage.
#上下文工程#AI编程#Anthropic#Claude
Then the model changed. Opus 5 and Fable 5 came out at half the cost and matched or exceeded the previous top tier (宝玉的分享, "Claude Opus 5 来了：半价摸到 Fable 5", 2026-07-25, https://mp.weixin.qq.com/s/). The old guardrails were not harmful in isolation. They had become harmful in aggregate — they competed with each other for the model's attention.
Thariq showed a typical trace from production: a single request where the system prompt said "leave documentation where appropriate" while a Skill said "never add comments". Both rules had been added for good reasons at different times. The model spent reasoning budget deciding which one to follow (Thariq, https://x.com/trq212/status/2080710971228918066).
The Anthropic team described it as five product managers standing behind one programmer, each saying one important thing, with the programmer burning cycles to guess who outranks whom (translated from 宝玉的分享, "Claude Code 的系统提示词，只剩不到 20% 了", 2026-07-26).
What replaced the deleted 80%
The team did not replace prompt mass with more prompt mass. They replaced it with three structural changes — and packaged them as the "Then / Now" ruleset for Claude 5 context engineering.
1. Stop teaching actions; design the interface instead
The old habit was to seed the system prompt with examples of how to use each tool. "Here's how to call the task list. Here's how to write a good commit." Examples lock the model into one exploration path. The Anthropic team found that Opus 5 reads a clean tool description and figures out the rest — if the description itself is clean.
The fix: spend engineering time on the shape of the tool. What does the parameter set express? What does the return value carry? If you can name the parameter clearly and constrain the return shape, the model infers usage. Thariq gave the example of a task-list tool: the moment you enumerate statuses as pending | in_progress | completed, you have already taught the model the state machine — without a single example sentence (Thariq, https://x.com/trq212/status/2080710971228918066).
The model does not need a tutorial. It needs a well-designed API.
2. Stop pre-loading everything; disclose on demand
The legacy instinct: dump every rule, every format, every edge case into the system prompt, because what if the model needs it later? Claude 5 has a wider context window, but attention is still a finite resource. Stuffing a hundred "might be useful" items into the same window as the three that actually matter makes the three harder to see.
The fix is progressive disclosure:
Core context: keep only what the current task genuinely needs.
Multi-step workflows (publishing, code review, verification): move them into Skills, loaded on demand.
The same principle applies to your own CLAUDE.md. Stop treating it as a central warehouse of every practice you've ever seen. Make it a tree: a thin top-level file plus deeper Skill files that load only when the agent enters that branch.
3. Stop hand-writing memory; hand it real references
For most of 2025, the standard recipe was: "every preference, every quirk, every past mistake — write it into CLAUDE.md." Anthropic's new guidance says no. Opus 5 already has automatic memory across sessions; it will surface the relevant context when it matters.
What CLAUDE.md should hold is only what the model cannot read from the filesystem: the build command, the one weird import rule, the team's forbidden library. Everything else — testing conventions, API patterns, design preferences — should be referenced from code, tests, HTML mockups, even an HTML page that shows what "good" looks like.
Thariq's rule of thumb: if you can express it as code or a test, the model learns it faster than if you write ten sentences about it (Thariq, https://x.com/trq212/status/2080710971228918066).
The Anthropic team also shipped a /doctor command inside Claude Code to help users audit and trim their own CLAUDE.md and Skills against these rules (Thariq, https://x.com/trq212/status/2080710971228918066).
What to delete from your own CLAUDE.md tomorrow
If you maintain a CLAUDE.md, Skills, or system prompt for any Claude-powered agent, the team offered a four-step audit. Treat it as a deletion checklist, not an addition checklist.
Keep only repo-specific quirks. The build command. The single-file-types rule. The library the team picked for a non-obvious reason. Everything a filesystem walk could tell the model, delete.
Move multi-step flows into Skills. Publishing, code review, verification. Each one is a Skill that loads only when the agent reaches that step.
Replace prose preferences with references. Don't write "API design should be clean." Point the agent at the cleanest file in your repo. Don't write "tests must cover edge cases." Show the agent a test that broke a real bug last week.
Cap CLAUDE.md at ~200 lines. Anthropic's own guidance: keep it under 200 lines. Beyond that, each additional rule loses adherence per-line (Thariq, https://x.com/trq212/status/2080710971228918066).
The first run of this audit will feel wrong. You will find rules you added in March 2025 that protected against a specific failure mode that no longer exists. Delete them.
What this means for the next model cycle
If the team's analysis generalizes — and the 80% result without eval loss is strong evidence that it does — then most teams writing agent harnesses today are over-specifying. They are writing Claude-3 instructions for Claude-5 models and paying for it twice: once in token cost on every request, once in degraded performance because the model has to resolve internal conflicts.
The teams that win the next 12 months are not the ones that ship the largest system prompt. They are the ones that ship the smallest system prompt that still holds — and pair it with well-shaped tools, lazily-loaded Skills, and a CLAUDE.md short enough to actually read.
The hardest skill for an agent builder in 2026 is not writing instructions. It is deleting them.
