The 6 month Claude cleanup

Why anyone would delete a working setup

Boris Cherny created Claude Code. Speaking at Y Combinator's Startup School this year, he told a room full of people to throw away the setup they'd carefully built.

He added that for Opus 5 he strongly recommends trying it, because the model may no longer need the long instructions that earlier models required. His own team practices this. When Opus 5 shipped, they cut more than 80 percent of Claude Code's built in instructions, and it got better.

Here's what that means for you. Every rule you added was written to fix something Claude was getting wrong at the time. "Stop being so wordy." "Always double check the math." "Use this exact format." Those were patches for a model that no longer exists. The current one often does that on its own, and your old patch now sits in front of it like furniture in a doorway.

My honest take, and this is where I part ways with the advice as it's usually repeated online: test before you delete. Deleting everything at once tells you something broke without telling you which part was carrying the weight. The checklist below finds that out in about 20 minutes per item.

Back it up before you touch anything

7 steps, one item at a time

Work through this for one skill or one instruction file at a time. Doing them all at once is how people end up reverting the whole thing in frustration.

1. Back it up. The step above. Don't skip it because you're only testing one small thing.
2. Write your list. Every skill, every project's instructions, every set of custom preferences. Most people are surprised here. A typical setup has 6 to 10 things quietly running, and 2 of them were built for a problem that stopped existing.
3. Pick the oldest one. Age is the best predictor. Something you wrote last month is probably still needed. Something from 8 months ago was written for a different model.
4. Run it bare. Open a brand new chat with no skill, no project, nothing attached. Give Claude a real task you'd normally use that skill for, and watch what comes back. Use the prompt below to make this comparison honest.
5. Judge the gap. If the bare result is as good, the skill has retired and you can delete it. If it's worse, the skill is doing real work. If it's close but missing your specifics, the skill needs 3 lines instead of 30.
6. Prune the survivors line by line. For anything you keep, delete one instruction and rerun the task. Nothing changed? That line was decoration. This is slower than deleting everything, and it's how you find out which parts were doing the work.
7. Book the next one. Put a repeating reminder on your calendar for 6 months out. Add a note about which items you retired, so future you isn't re-testing the same things.

Make Claude audit itself

Testing by feel doesn't work, because you'll read the answer you expect. This prompt makes Claude do the same task twice, with and without your instructions, then tell you exactly which lines earned their place. Paste it into a fresh chat with nothing else attached.

I want to audit one piece of my Claude setup and find out whether it still earns its place.

Here are the instructions I want to test:
[paste the full text of your skill or your instructions here]

Here is a real task I would normally use them for:
[describe the task in a sentence or two, and paste any example material you would normally give it]

Do this in 4 steps:

1. Complete the task now using only what you already know. Completely ignore the instructions I pasted above.

2. Complete the task a second time, this time following those instructions exactly.

3. Compare the two versions. Tell me specifically what the instructions changed. Quote the lines that made a real difference to the output, and list the lines that changed nothing at all.

4. Give me a verdict: keep all of it, keep only certain lines (write out the trimmed version), or retire it because you handle this well without it.

Be blunt with me. If the instructions made no real difference, say so plainly instead of being polite about it.

Run this on your oldest skill first. The trimmed version it writes in step 4 is usually the one you want to keep.

What usually goes, what usually stays

After doing this on my own setup, the pattern is consistent. Instructions about how Claude should behave tend to expire. Instructions about your actual business never do.

That split is the useful part of this whole exercise. What survives a cleanup is the part that's actually yours, and a shorter setup made only of that runs better than a long one carrying 6 months of patches.

Two things worth knowing