Hey HN. I use Claude Code in auto mode a lot and kept catching myself not reading the commands it runs. They scroll past fast, and half of them are find | xargs | awk one-liners I'd need a minute to parse anyway
So I made a small plugin that puts one plain line above the prompt while the call is running, like "Survey disk usage by file type", and for longer pipelines it spells out the steps. If the command looks destructive (rm -rf, force push, curl | sh, that kind of thing) the line goes red and you get a warning for a few seconds
It works off the tool's own arguments, so by default there's no extra model call and nothing leaves your machine. There's an optional mode where it asks Claude to reword the unclear ones, mostly edits
The risk check is just pattern matching on the command text, so think of it as a heads-up, not a safety net
Hey HN. I use Claude Code in auto mode a lot and kept catching myself not reading the commands it runs. They scroll past fast, and half of them are find | xargs | awk one-liners I'd need a minute to parse anyway
So I made a small plugin that puts one plain line above the prompt while the call is running, like "Survey disk usage by file type", and for longer pipelines it spells out the steps. If the command looks destructive (rm -rf, force push, curl | sh, that kind of thing) the line goes red and you get a warning for a few seconds
It works off the tool's own arguments, so by default there's no extra model call and nothing leaves your machine. There's an optional mode where it asks Claude to reword the unclear ones, mostly edits
The risk check is just pattern matching on the command text, so think of it as a heads-up, not a safety net
Curious what else you'd want it to catch