← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

The Context
Cleanup Checklist

Anthropic cut over 80% of Claude Code’s system prompt and its coding scores didn’t drop. Here’s the checklist I use to trim my own AI setup so it thinks better, not harder.

You commented TRIM, so here it is. In July, Anthropic published new context engineering rules for its Claude 5 models. In it, they say they removed over 80% of Claude Code’s system prompt for models like Opus 5 and Fable 5 “with no measurable loss” on their coding evaluations.

The lesson for the rest of us: most AI setups aren’t underfed, they’re overloaded. Fewer, sharper instructions beat a wall of rules. This is the pass I run on my own setup, and it takes about ten minutes.

The Checklist
  1. Delete every rule that states the obvious. “Write good code,” “be helpful,” “use correct grammar.” Modern models already do this, and every line you cut is one less thing competing for attention.
  2. Turn rules into outcomes. Anthropic swapped a strict “no comments, one short line max” rule for one line telling Claude to write code that reads like the surrounding code, matching how it comments and names things. Say the result you want and let the model work out the details.
  3. Kill the conflicts. Anthropic found requests where the system prompt, skills and user said things like “leave documentation as appropriate” and “DO NOT add comments” at the same time. When two lines fight, the AI has to guess. Pick one.
  4. Break the giant file into small ones. Keep the always-true stuff short and load the rest only when the task needs it. Anthropic calls this progressive disclosure. In Claude Code, that means a verification skill your CLAUDE.md points to. In the Claude app, it means a skill or a Project file.
  5. Show the real thing. Anthropic says an HTML mockup of a design usually beats a written description or a screenshot. Point Claude at the actual file, template or spec. Go easy on piles of examples, though. Anthropic found they can box the model into copying them.
  6. Keep the gotchas. Spend your words on the surprising stuff Claude would get wrong without being told. Skip anything it can see for itself by looking at your files.
The Cleanup Prompt

Prefer to skip the manual audit? Paste your custom instructions, Project instructions or CLAUDE.md into Claude with this prompt. It hands back a trimmed version plus a list of what it cut.

The Context Cleanup Prompt
You are my context editor. I'm going to paste my current AI instructions below. Your job is to make them shorter and sharper without losing anything that actually changes how you behave. Apply these rules: 1. Delete anything that states the obvious, that a capable model already does by default, or that you could figure out by looking at my files. 2. Rewrite rigid rules as the OUTCOME I actually want. For example, "write code that reads like the surrounding code" instead of ten formatting rules. 3. Find instructions that conflict with each other and tell me how you would resolve each one. 4. Separate the always-true instructions (keep these short) from task-specific ones that should live in their own skill, project file or doc and load only when relevant. 5. Point out where a real reference (an actual file, template, HTML mockup or spec) would say it better than a long description. Flag any pile of examples that might box you in. 6. Keep the non-obvious gotchas I'd get wrong if you weren't told. Those deserve the most words. Return three things: A) The trimmed instructions, ready to paste back in. B) A short list of what you cut or moved, and why. C) Anything you were unsure about, so I can make the call. Don't invent new rules. If something is unclear, ask me instead of guessing. Here are my current instructions: [PASTE YOUR CUSTOM INSTRUCTIONS, PROJECT INSTRUCTIONS, OR CLAUDE.MD HERE]

The Real Win

You’re not chasing the perfect prompt anymore. You’re curating context. Give AI less to wade through and clearer references, and it gets sharper on its own.

Want the bigger picture? I broke down the context engineer role here. And when a chat gets bloated mid-task, my Save Your Claude Context handoff resets it clean.