When in doubt, just dump it in auto-memory

Touching Claude Code, the first thing you hit is the question:

…what goes where, exactly?

Sure, designing it properly upfront pays off. Organizing repeated instructions and prohibitions makes AI output a lot more stable.

But honestly, if you try to design everything cleanly from day one, the configuration alone eats a whole day.

So lately, I think the right level is: when in doubt, “just throw it in auto-memory” and move on.

(From here, “memory” means auto-memory.)

Claude Code is sharp enough that even when you casually hand it stuff like

it actually picks it up surprisingly well.

If you want something remembered for sure, use an explicit phrase like “record this rule to memory.”

This is genuinely useful.

Of course, there’s a side effect.

Later you end up wondering “why is it doing this?”

The situation naturally shifts from when you first had it remember something, but whether memory updates itself to match is a matter of luck.

Settings get scattered, and AI’s “implicit knowledge” grows.

The trickier part: it’s hard to share via Git.

CLAUDE.md or harness configs can sit in the repo, but anything in memory stays in your personal environment.

Which means from another teammate’s or another AI’s perspective,

“why did it make that judgment?” becomes invisible.

The AI looks like it’s getting smarter on its own, but really, local context dependence is just piling up.

So:

It’s a trade-off.

Personally, rather than trying to build the perfect AI-development setup from day one, the realistic flow is:

Studying tools to boost efficiency is undeniably important.

But it’s also dangerous to end up only researching tools and never shipping anything.

This happens especially easily in AI circles.

Once you start chasing “latest setup” “strongest harness” “full automation”

you spend forever on environment setup.

That’s why “knowing better and choosing to cut corners” matters a lot.