The instinct, when an AI coding agent gets something wrong, is to feed it more: a longer instructions file, more rules, more caveats. The people closest to these tools keep giving the opposite advice, and after shipping real work with an agent every day, we think they are right.

More context is not more control

A giant instructions file feels like control, but it usually reads to the model as noise. Every extra rule competes for attention with the actual task, and contradictory or stale instructions quietly drag quality down. The uncomfortable truth is that a bloated context file is often the reason your agent feels dumber this month than last, you have been training it on your own clutter.

What we keep, and what we cut

Our own working file is short on purpose: the handful of conventions that are genuinely non-obvious and genuinely enforced, and nothing else. If a rule is already obvious from the code, it goes. If it is aspirational rather than real, it goes. What survives is the small set of things a sharp new teammate would actually need told to them, which, not coincidentally, is the same test we apply to whether a role should exist at all.

Why this matters beyond code

The lesson generalizes to every AI system we build for clients: the quality lever is rarely “add more instructions,” it is “remove the ones that no longer earn their place.” A lean, current context beats an exhaustive, rotting one every time, the same discipline that keeps automation debt from piling up. If you want that applied to a real workflow, that is what our solutions are.

Sparked by “Claude Code’s Creator: Delete Your CLAUDE.md, 7 Moments That Matter (Boris Cherny at YC)”. Credit to the original, watch it for the full argument; the take above is ours.