9 checks for your instructions file

1. What happened

On July 24, 2026, Anthropic announced it had removed more than 80% of Claude Code's system prompt for the new models, with no measurable drop in code test results. The team that wrote that prompt deleted most of it, and published the rules it used to decide what stayed and what went.

• Most of us do the exact opposite: every agent mistake turns into one more line in the instructions file.
• The file only grows, nobody deletes anything, and the model keeps improving in the meantime.
• The result: you're paying context on rules written for a model that no longer exists.

2. The nine checks

Go through the file line by line. Any line that fails one of these checks is a candidate for deletion or for moving elsewhere.

• Who was this line written for? If it was written a model generation ago, there's a good chance it's no longer needed.
• Is this a prohibition or a judgment call? "Write like the surrounding code" beats ten lines of don't-do-this.
• Is this a procedure or a fact? A long process is a skill, not a line in the file.
• Does this instruction already appear somewhere else? The new model picks it up from one mention.
• Is this something every developer already knows? The file should keep the repo's specific traps, not general knowledge.
• Can you replace the description with code? An existing test or a similar file explains more than a paragraph.
• Does the example still work for you? A clear tool interface teaches its own usage without examples.
• What does it cost? Check how much space the skills and the file take up before you add more.
• What's left? A short file that holds only what's surprising about this particular repo.

3. The ready-made prompt

Open a session in Claude Code inside the repo, and paste:

Go through this repo's instructions file (CLAUDE.md or AGENTS.md, including nested files).
  
  For every line or section, answer the nine questions:
  1. Was it written for an old model generation?
  2. Is it a prohibition that could be a judgment call instead?
  3. Is it a procedure that should move into a skill?
  4. Is it a duplicate of an instruction that appears elsewhere?
  5. Is it general knowledge every developer already has?
  6. Could the description be replaced by a reference to existing code, a test, or a file?
  7. Is it an example that's no longer needed?
  8. What does it cost in context relative to its benefit?
  9. If it stays, what specific trap does it prevent?
  
  Return a table with: the line, the recommendation (keep / shorten / move to skill / delete), and the reasoning.
  
  Don't edit any file. I want to see the proposal first.
  At the end, write separately which sections you'd pull into a skill and what name you'd give it.

4. What not to delete

Shortening isn't the goal. An instruction that prevents a repeat mistake is worth more than the tokens it costs.

• Traps specific to this repo: a non-standard build command, a service that needs to run first, a file that looks redundant but must not be deleted.
• Decisions you already paid for with a bug. That's exactly what the file is meant to remember.
• Permission boundaries: what the agent is allowed to do on its own and what needs your approval.
• A skill you pulled out of the file still needs to be good. Moving it isn't deleting it.

5. How to check you succeeded

After the cut, give the agent a real task that succeeded before, and check that it still succeeds. If something breaks, bring back one line, not the whole file. There's a separate guide for measuring the context itself: /context shows what's taking up the window.

• Test on a task where you already know what the result should look like, not a new one.
• Bringing back one line at a time teaches you what actually held the behavior in place.
• A line you've brought back twice is a good candidate to become a named skill.

6. Sources

The rules in this guide are based on Anthropic's publication and Claude Code's official documentation.

https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
  https://code.claude.com/docs/en/skills
  https://code.claude.com/docs/en/memory
  https://code.claude.com/docs/en/context-window