# Claude 5 Context Engineering: Anthropic Deleted 80% Prompt > Source: > Published: 2026-07-26 01:11:57+00:00 On July 24, 2026 — the same day Claude Opus 5 launched — Anthropic quietly published something more useful than a model announcement: proof that its own system prompt was 80% dead weight. Thariq Shihipar, a Member of Technical Staff at Anthropic, revealed that the team deleted over 80% of Claude Code’s system prompt when migrating to Claude Opus 5 and Fable 5, with no measurable loss on coding benchmarks. The lesson landed hard: the model got smarter, the prompts didn’t catch up. If Anthropic’s own system prompt was bloated with unnecessary guardrails, yours almost certainly is too. Every CLAUDE.md file built for Claude 3 behavior carries the same legacy risk — instructions that once tamed a less capable model now create conflicting guidance for a more capable one. ## What “Unhobbling” Actually Means Shihipar’s term for the deletion is blunt and accurate: “unhobbling.” The fix wasn’t better few-shots or more precise wording — it was removing constraints designed to prevent failures that Claude 5 simply no longer produces. According to [Anthropic’s official context engineering guide](https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models), the old prompt contained conflicting messages that forced the model to deliberate unnecessarily — the exact opposite of what a capable model needs. The concrete example makes this tangible. The old Claude Code instruction read: “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 observe a codebase’s commenting style and match it without being told what “matching” looks like. This isn’t a minor tweak. It’s a signal that the standard playbook for prompting — be explicit, be exhaustive, cover every edge case — actively degrades performance on Claude 5 models. The model outgrew the scaffolding. ## The Four Shifts You Need to Make Shihipar’s guidance distills the change into four concrete transitions. First: rules to judgment. Stop writing “never do X” instructions for behaviors Claude 5 already handles well. Trust the model’s judgment for decisions it can make from context, and reserve explicit rules for behavior the model genuinely cannot infer. Second: examples to interface design. Providing tool usage examples constrains exploration. [Shihipar explains](https://the-decoder.com/anthropic-says-it-cut-80-percent-of-claude-codes-system-prompt-because-fable-5-models-want-a-smaller-system-prompt/) that 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 more clearly than any example — without narrowing the model’s problem-solving space. Third: everything upfront to progressive disclosure. CLAUDE.md shouldn’t be a dump of everything the model might ever need. Move verification workflows and review checklists into Skills that Claude calls selectively. Load heavy context when it’s relevant, not every session. This directly reduces the context budget consumed before Claude even starts working. Fourth: repetition to 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. Redundancy isn’t clarity — for frontier models, it’s noise. ## The Context Budget Problem Here’s the math that makes this urgent. Frontier models reliably follow roughly 150–200 instructions. Claude Code’s built-in system prompt already consumes approximately 50 of those slots. Every instruction in your CLAUDE.md competes directly with Claude Code’s own operational guidance. A 200-line CLAUDE.md isn’t giving Claude 200 instructions — it’s overloading a constrained instruction set and generating conflicts. Anthropic built a tool to address exactly this: the `/doctor` command in Claude Code automatically identifies instructions that are no longer needed for newer models and suggests simplifications. Run it. The suggestions are based on the same analysis Shihipar’s team applied when deleting their own prompt. According to [post-launch analysis from the developer community](https://www.mager.co/blog/2026-07-24-context-engineering-claude-5/), the migration is largely subtraction — “pull every ‘verify,’ ‘double-check,’ and ‘re-verify’ instruction out of your prompts.” ## What Actually Stays The deletion isn’t a directive to gut everything. What survives is context the model cannot derive from reading the codebase: brand voice requirements, pricing constraints, sign-off authority, confidentiality rules, and non-obvious architectural decisions specific to your product. If Claude could infer it by reading the surrounding code, delete it. If it requires business knowledge the model can’t possibly have, keep it. The distinction matters because overcorrecting is a real risk. Developers who delete everything may strip the genuine irreducibles — the rules that aren’t about model capability at all, but about business requirements. Context engineering for Claude 5 isn’t “less is more.” It’s “only what’s necessary.” That’s a different — and more demanding — standard. ## Key Takeaways - Anthropic deleted 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no performance loss — if their prompt had dead weight, yours does too. - The four shifts: rules to judgment, examples to interface design, everything upfront to progressive disclosure, repetition to single-source instructions. - Frontier models follow roughly 150–200 instructions maximum; Claude Code uses approximately 50, leaving far less room for CLAUDE.md than most developers assume. - Run `/doctor` in Claude Code to automatically identify instructions no longer needed for Claude 5 models. - Keep only what Claude cannot infer: business constraints, brand rules, confidentiality requirements, and genuinely non-obvious architecture decisions. Delete the rest.