The best part of your Claude Code is what you delete

I’m writing a new package currently called anvil-skills, it’s a skillset based on what I already use in production.

How do I fix 2 big Claude problems:

1. rules ignored mid-session.

2. learnings that only apply to one kind of task.

The more important part is the 2nd problem.

We all know this tax, add something to CLAUDE.md and well it is now running on every call on every session.

This isn’t just about keeping everything short, that’s a lot of generic advice that leads to crap results.

My first design idea, was making these reads conditional.

For example I have /scope which has no idea about REVIEW_RULES.md.

But the learnings are still captured. More importantly the feedback is logged.

The transferable thing of this setup, is it applies to all skills.

This is something I’ve been fighting personally and we have been fighting at work.

Stale is easy for Claude to assume and then it spends time finding a correction, it will also point that out to you and ask if it should update so that’s good.

However it’s also sending Claude on a longer path than it would otherwise.

So I like to sit in the middle if possible.

Stale business logic and docs are fine, generic Claude stale recommendations and commands are useless.

It’s a hard battle.

This is more specific to my package but applies for a team, and you across all your projects.

I nearly shipped "slice so one ticket is one PR a human can review in one sitting".

However that’s hyper specific right?

For work that’s awesome! For my game that sucks, for my personal website awful.

Sometimes empty files with titles are fine, just ignore Claude always wanting to delete it.

This has been a massive learning from the package, you should use now for your process.

If you add a rule to CLAUDE.md or your own SKILL or a context file.

Audit everything else and remove stale or dead items.

This leads to cleanup that never happens.

People say run this audit every x amount of time but who really does that? Not me.

I like to get claude to do this. Automatically via a skill or when updating something.

It’s hard for us to constantly track learnings, now more than ever with all the projects we have.

So get Claude to help you with it.

Set up a feedback log.

Then find learnings in those logs.

Simple as that.

Check out my repo: https://github.com/Alexanderdunlop/anvil-skills

Stay tuned it’s going to be awesome