Claude 5 Context Engineering: Cutting 80% of System Prompts Without Performance Loss
Anthropic reveals how they removed over 80% of Claude Code's system prompt for Opus 5 and Fable 5 models with no measurable benchmark regression, replacing rigid constraints with progressive disclosure, dynamic skills, and trust in the model's improved judgment.
Unbinding Claude
Anthropic discovered that both system prompts and CLAUDE.md files had accumulated excessive constraints for Claude Code. Internal logs showed conflicting instructions within single requests — such as "retain documentation judiciously" versus "do not add comments" — because system prompts, skills, and user requests clashed. While Claude could still infer user intent, it had to reason through overlapping, contradictory guidance before acting.
Many of these guardrails were originally necessary to avoid worst-case behaviors in earlier models. However, the team found they could delete most constraints and let the model rely on surrounding context and its own judgment. For Claude Opus 5 and Claude Fable 5, they removed over 80% of the system prompt content with no measurable loss in coding benchmarks. Additionally, Claude Code now has more tools: memories, Artifacts, and Skills provide new ways to load and share context across sessions, reducing the need for everything to live in CLAUDE.md.
Past vs. Present
Previous context-engineering best practices have become myths. One example: a rule stating "do not write comments by default; never write multi-line docstrings or comment blocks — at most one short line; do not create planning, decision, or analysis documents unless asked." This guidance fails in documentation scenarios where users have their own preferences, or in complex code sections that genuinely need multi-line comments. Without these guardrails, older models often produced incorrect comments, a trade-off the team accepted. New models exhibit better judgment and handle such decisions well even without explicit rules.
The new system prompt instead says: "When writing code, make it read like the surrounding code: match its comment density, naming style, and idiomatic patterns."
Rather than relying on examples, the team now focuses on tool, script, and file design — what parameters Claude has and how to make them more expressive. In the Todo tool, simply enumerating statuses as pending, in_progress, and completed cues Claude on usage; an instruction to keep one item in_progress defines the expected behavior.
Because Claude Code specializes in coding, the system prompt included detailed code-review and verification guidance. These are not always needed but are critical when they are. The team moved verification and code review into separate skills that Claude Code can invoke selectively — an instance of progressive disclosure: loading the right context at the right time.
Progressive disclosure also applies to tools. Some tools are "lazy-loaded": the agent must first search for their full definition via ToolSearch before using them. This allows more tools (e.g., the Task tool) without consuming context until needed.
The same approach works for personal CLAUDE.md and Skill.md files. A common pitfall is creating a central repository of every known practice you might encounter, fearing Claude won't find them otherwise. Instead, consider a file tree that loads relevant files at the appropriate time (see https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code).
Early Claude models sometimes needed repeated instructions or favored instructions at the end of the context window. This led to duplicating tool references in both the main system prompt and tool descriptions. The team removed these redundancies, placing tool-usage instructions in the tool descriptions themselves.
Users were previously encouraged to save content to Claude's memory via the # hotkey to auto-write CLAUDE.md (http://claude.md/). Now Claude automatically saves memories relevant to your work and to you personally.
In planning mode, Claude Code heavily relied on Markdown files containing plans. Storing plans as files helped Claude reference them when needed. A similar practice was keeping specifications in the codebase for long-running projects. However, Claude can now handle increasingly complex references: not just simple Markdown files, but also HTML Artifacts created by the new Artifacts feature.
References can also be provided as code: a detailed test suite, or a function from another codebase that Claude may need to port. Rubrics are another reference form — they let Claude verify your taste in a specific domain (e.g., what good API design looks like) by using dynamic workflows (https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code) and launching verification agents with those rubrics.
Applying This to Your Context
Use progressive disclosure to surface more detail. If you have unique verification instructions, create a verification skill and reference it in CLAUDE.md. For longer skills, apply progressive disclosure aggressively — split them into multiple separated files.
Skills work best when they encode opinions, knowledge, or best practices specific to you, your team, or your product. These may live in spec files, prototypes, or entire codebases. Prefer code-form files because they give Claude clear, high-fidelity instructions in a language it knows well. For example, an HTML prototype of a design typically yields better results than a design description or screenshot.
Try Simplifying
You may need to simplify your own system prompts, skills, and CLAUDE.md files, just as Anthropic did. A new command claude doctor (run /doctor in Claude Code) helps automate this. For more details on prompting advanced models, see the Fable field guide (https://claude.com/blog/a-field-guide-to-claude-fable-finding-your-unknowns).
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
