The deviation log is the most underrated pattern here. When you're running agents against a plan, drift is inevitable — what matters is having a record of where it drifted and why. I do something similar with Claude Code: keep a PLAN.md and have the agent update it when scope changes. Saves hours during review.
The "rage-quit prevention" threshold for shipping is also real. If users find enough value to request features, you've already validated the core. Ship the skeleton, let real usage tell you what to add next. Most solo builders (myself included) over-polish before anyone's touched the product.
I haven't found the right prompt yet, but I've been thinking about how to create a deviations log for Kilo itself. Every time there's a process fail - a tool call, api timeout, task id missing, ..., anything that adds friction to what we're actually trying to do - add log entry to papercuts.md.
I'm not sure this should in the standard context itself because of the potential to derail the main task into recursive never-ending sidequests. Maybe a tool to inspect and review threads after the fact. Feels like something that could/should be baked into Kilo.
The deviation log is the most underrated pattern here. When you're running agents against a plan, drift is inevitable — what matters is having a record of where it drifted and why. I do something similar with Claude Code: keep a PLAN.md and have the agent update it when scope changes. Saves hours during review.
The "rage-quit prevention" threshold for shipping is also real. If users find enough value to request features, you've already validated the core. Ship the skeleton, let real usage tell you what to add next. Most solo builders (myself included) over-polish before anyone's touched the product.
I haven't found the right prompt yet, but I've been thinking about how to create a deviations log for Kilo itself. Every time there's a process fail - a tool call, api timeout, task id missing, ..., anything that adds friction to what we're actually trying to do - add log entry to papercuts.md.
I'm not sure this should in the standard context itself because of the potential to derail the main task into recursive never-ending sidequests. Maybe a tool to inspect and review threads after the fact. Feels like something that could/should be baked into Kilo.