Anthropic have published the new rules of context engineering for the Claude 5 generation. The headline finding is a bit humbling for anyone who has spent months building elaborate instruction files: they cut over 80% of Claude Code's system prompt and their coding evals did not drop.
Your prompt is only a sliver of what the model actually receives. The rest gets assembled from system prompts, skills, CLAUDE.md files, memory and whatever else you have wired in. That assembly job is context engineering, and the rules for doing it well have changed.
The five shifts
- Rules become judgement. Blanket commands like "never write comments" were guardrails for weaker models. They are wrong a fair slice of the time, and newer models read the surrounding context and make a better call themselves. Describe the outcome, not the prohibition.
- Examples become interface design. Showing Claude how to use a tool used to be rule number one. Now examples quietly fence it into the narrow space you demonstrated. Spend that effort on the tool itself: clearer parameters, tighter option sets, names that explain themselves.
- Everything upfront becomes progressive disclosure. Stop treating your instruction file as a warehouse for everything you might one day need. Split it into a tree of smaller files that load only when relevant. Tools can work the same way, sitting unloaded until something calls for them.
- Repetition becomes simple descriptions. Older models needed the same instruction in three places to be sure it landed. Put guidance about a tool inside that tool's description, then delete the copies.
- Simple specs become rich references. A spec no longer has to be a markdown doc. A test suite is a spec. A working function in another codebase is a spec. An HTML mockup beats a written description of a design. Rubrics let you encode your taste so Claude can verify its own work against it.
One more: memory now saves what is relevant automatically, so hand-maintaining your own memory file is largely a job you can stop doing.
Where everything belongs
- System prompt: tied to the product. If you are building your own agent, this is where the time goes.
- CLAUDE.md: keep it light. A line on what the repo is for, then most of the tokens spent on genuine gotchas, the odd conventions nobody could guess from reading the code.
- Skills: your opinions, your team's practices, the things specific to you. Split long ones across files rather than one monster.
- References: the detail. Prefer things written in code, because that is a language Claude already knows well.
The practical move
Run /doctor in Claude Code. Anthropic have baked these practices into a command that audits your skills and CLAUDE.md files and tells you what to cut. Then read your instruction file and your skills side by side and look for places they contradict each other, the classic being one asking for thorough documentation while another bans comments. Every clash is work the model does instead of your task.
Read the full post from Anthropic