Problems people are actually hitting.
One problem per post, what causes it, and what to do. No listicles, and nothing here is selling you the fix — most of these are settings and habits, not products.
Auto-compact is eating your session
4 min“Claude was doing fine, then it forgot a decision we made an hour ago.”
Claude Code summarises your conversation when the context fills up. The summary is lossy by design, which is why the model forgets a constraint you set an hour ago. Compact at task boundaries instead.
Why subagents produce shocking bills
4 min“One command ran for two hours and the bill was four figures.”
Reports of five-figure Claude Code sessions have one shape in common: parallel subagents, each with its own context, and nothing that stops them. Here is the mechanism and the ceiling to put on it.
Your prompt cache expires faster than you think
3 min“Same project, same prompts, and the bill went up for no reason I can see.”
Cache lifetimes are measured in minutes, not hours. If you think between prompts, you are paying full price for the same prefix over and over — and the fix is about rhythm, not configuration.
A GitHub issue is untrusted input
4 min“Our CI runs an agent on new issues. Someone pointed out that anyone can open an issue.”
Two 2026 advisories describe the same shape: an agent reads attacker-controlled text and treats it as instructions. If your agent can read the internet and reach a secret, that is the whole vulnerability.
Reading the docs is not studying
4 min“I read everything and still did not feel ready in the room.”
The common report from Claude certification candidates is that they prepared by reading and were tested on judgement. Retake waits make that an expensive way to find out. Here is a cheaper order to do it in.