A practical guide to updating your system prompts, CLAUDE.md files, skills, and agent harnesses for the latest generation of Claude models.

When you send a message to Claude, the prompt you type is only a small part of the context the model actually receives. The rest is assembled from the system prompt, Skills, CLAUDE.md files, memory, tools, and other sources. We call this context engineering, and it has a much larger impact on results than most people realize — especially when using Claude Code or building your own agents.

Unlike a one-off prompt, context is reused across many requests. It cannot be overly specific to a single task. The challenge is writing general guidance that still works well no matter what the user asks next. This becomes even harder as Claude’s capabilities improve.

With the newest generation of Claude models (Claude Opus 5, Claude Fable 5, and related releases), Anthropic discovered something surprising: they could remove more than 80% of Claude Code’s system prompt with no measurable drop in coding performance. The old rules of context engineering had become outdated. Here’s what changed and how you should update your own context.

Unhobbling Claude: Less Constraint, More Judgment

The biggest shift is simple: previous versions of Claude needed strong constraints to avoid worst-case behavior. Newer models do not.

When Anthropic reviewed internal transcripts, they found conflicting instructions appearing in the same request — for example, one part of the context saying “leave documentation as appropriate” while another said “DO NOT add comments.” Claude could usually resolve the conflict, but only after extra thinking.

Many of those hard rules were written to prevent older models from doing damage (deleting files, writing bad comments, creating unnecessary documents). With the new models, those same rules often create friction instead of safety. The better approach is to remove the rigid constraints and let Claude use surrounding context and its own improved judgment.

Claude Code also now has more powerful tools for memory, artifacts, and skills. What used to live only in CLAUDE.md can now be handled more dynamically.

Then vs Now: The New Rules of Context Engineering

Here are the key shifts Anthropic observed:

Then: Give Claude rules Now: Let Claude use judgment

Old system prompts contained strong, absolute instructions such as:

“In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max. Don’t create planning, decision, or analysis documents unless the user asks for them.”

These rules prevented bad behavior on older models, but they were frequently wrong for real user requests. Newer models handle these decisions well on their own. The updated guidance is much lighter:

“Write code that reads like the surrounding code: match its comment density, naming, and idiom.”

Then: Give Claude examples Now: Design better interfaces

The old number-one rule for tools was “give examples.” With current models, examples can actually limit exploration. Instead of showing Claude how to use a tool, focus on designing the tool’s interface so the correct usage is obvious from the parameters themselves.

Example: a Todo tool that exposes status as an enum (pending, in_progress, completed) and includes a note that only one item should be in_progress at a time teaches the model the intended behavior without needing long example blocks.

Then: Put everything upfront Now: Use progressive disclosure

Earlier system prompts packed detailed instructions about code review, verification, and other workflows into the main context even when they were not always needed. Claude Code now relies heavily on progressive disclosure — loading the right information only when it is relevant.

Verification and code review were moved into dedicated skills that Claude can call selectively. Some tools are “deferred loading”: the agent must search for their full definition with ToolSearch before using them. The same principle applies to your own CLAUDE.md and Skill.md files. Instead of one giant central file, create a tree of files that can be loaded at the right time.

Then: Repeat yourself Now: Simple tool descriptions

Older models sometimes needed the same instruction repeated in multiple places, or paid more attention to the end of the context window. Many repeated tool references have been removed from the system prompt. Instructions now live cleanly in the tool descriptions themselves.

Then: Memory in CLAUDE.md files Now: Auto-memory

Users were previously encouraged to save important information to CLAUDE.md with the # hotkey. Claude now automatically saves relevant memories about the work and the user.

Then: Simple specs Now: Rich references

Plans and specs used to live in simple markdown files. Claude can now handle much richer references: HTML artifacts, detailed test suites, functions from other codebases, or rubrics that define what “good” looks like in a particular domain. These higher-fidelity references produce better results than plain text descriptions.

How to Apply This to Your Own Context

System Prompt This is tightly tied to the product. It tells Claude what product it is operating in and what its role is. If you are building your own agent harness, invest heavily here. For Claude Code users, you will rarely need to change it.

CLAUDE.md Keep it lightweight. Briefly describe what the repository is for, then focus most of the tokens on gotchas that Claude cannot easily discover by reading the codebase. Avoid restating the obvious. Use progressive disclosure: if you have detailed verification steps, put them in a skill and reference the skill from CLAUDE.md.

Skills Treat skills as lightweight guides that Claude can discover when needed. Avoid over-constraining them except in high-stakes areas. For long skills, split content across multiple files and use progressive disclosure. The best skills encode opinions, knowledge, or practices that are specific to you, your team, or your product.

References You can @ mention files to include them as references. Prefer high-fidelity sources: code, test suites, HTML mockups, or detailed artifacts rather than vague descriptions or screenshots. Claude understands code extremely well, so a working example or a test suite often works better than a written specification.

Try Simplifying

Across your system prompt, skills, and CLAUDE.md files, the most valuable action is often deletion. Anthropic removed more than 80% of the previous Claude Code system prompt with no loss in performance. Over-constraining the model is now more harmful than under-constraining it.

Claude Code includes a new command, claude doctor (use /doctor), that helps right-size your skills and CLAUDE.md files automatically. For more detailed guidance on prompting the newest models, see Anthropic’s Fable field guide.

Final Takeaway

Context engineering for Claude 5-class models is less about writing more rules and more about creating clean interfaces, progressive disclosure, and high-quality references while trusting the model’s improved judgment.

The old best practices were designed for a less capable model. The new best practice is simpler: remove unnecessary constraints, design better tools and files, and let Claude do more of the thinking.