Anthropic found that with the release of stronger reasoning models like Claude Opus 5, old prompt engineering paradigms may be outdated. The team removed over 80% of Claude Code's system prompts without measurable performance loss on coding benchmarks. This suggests that as model capabilities improve, fewer constraints are needed, allowing models to rely on their own judgment.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
With the launch of new strong reasoning models like Claude Opus 5, Anthropic finds that old prompt engineering paradigms may be outdated or even counterproductive. They removed over 80% of system prompts in Claude Code with no measurable performance loss on coding benchmarks. This marks a shift toward a new paradigm of 'context engineering.'
Anthropic team member Thariq Shihipar posted on the launch day of Claude Opus 5, discussing the engineering trend changes brought by the new generation of models. Anthropic found that for models like Claude Opus 5 and Claude Fable 5, the team removed over 80% of the content from Claude Code's system prompts, with no measurable performance loss on coding benchmarks.
Anthropic points out that when users send a message to Claude, the prompt is only a small part of the context the model receives. The vast majority of context comes from system prompts, Skills, CLAUDE.md files, memory, and other sources—this is called 'context engineering.' Whether using Claude Code or building your own agent, it significantly impacts the quality of the final output.
Anthropic found that some past 'best practices' have become pitfalls. For example, old system prompts included rigid instructions like 'by default, don't write comments in code' and 'never write multi-paragraph docstrings.' But new models have better judgment and can handle these without explicit rules. The new approach becomes: 'Write code that reads like the code around it: match its comment density, naming conventions, and idiomatic style.'
Another pitfall is providing Claude with numerous usage examples. Anthropic found that examples can actually constrain the model's exploration space. Instead of piling on examples, it's better to invest effort in designing tools, scripts, and files to make parameters more expressive. Additionally, the old practice of repeatedly emphasizing instructions is no longer necessary; tool usage instructions should be placed in tool descriptions rather than system prompts.
Anthropic emphasizes 'progressive disclosure': loading the right context at the right time. For instance, they split validation and code review into separate Skills that Claude Code can call on demand. Some tools use 'lazy loading,' where the agent must first retrieve the full definition via ToolSearch before calling them.
CLAUDE.md should remain lightweight, briefly describing the repository's purpose and focusing on documenting 'gotchas' in the codebase. High-performing teams keep it under 60 lines, and it should typically never exceed 300 lines. Skills are treated as lightweight guides; avoid over-constraining unless absolutely critical. For longer Skills, split them into multiple files for divide-and-conquer.
In plan mode, Claude Code used to rely heavily on markdown-formatted plan files. But Anthropic found that Claude can now handle more complex reference materials, such as HTML artifacts and code itself. A specification can be a comprehensive set of test cases, or a function to be ported from another codebase. Rubrics are also a type of reference material; Claude can derive a validator agent through dynamic workflow to check developer taste.
Anthropic recommends prioritizing files in code form as reference materials, as they provide clear, high-fidelity instructions to Claude. For example, an HTML design mockup often works better than a textual description or a screenshot of the design.
In the past, users were encouraged to use the # shortcut to automatically write content into CLAUDE.md to save memory. Now, Claude automatically saves memories related to the current work and user. Additionally, Claude Code now has more tools, such as memory, artifacts, and Skills, creating new ways to load and share context across sessions.
Anthropic also introduced the claude doctor command. Using /doctor in Claude Code automatically adjusts Skills and CLAUDE.md files to appropriate sizes, helping developers simplify.
This article's information primarily comes from public posts by Anthropic team member Thariq Shihipar and reporting by Machine Heart. All specific numbers (e.g., 80%, 60 lines, 300 lines) originate from that reporting and have not been independently verified by third parties.
The stronger the model, the less it needs lengthy rules and examples. Prompt engineering is shifting to context engineering, with a core focus on simplifying system prompts, progressive disclosure, leveraging Skills and reference materials, and letting the model judge for itself.
Primary report
Primary source
Also covered by 1 sources