context-reflect

Invoke context-reflect immediately after the task completes while the conversation and tool-call history are still visible. The skill reconstructs phases of work, skills used, files touched, and validation performed, then outputs a structured retrospective report.

The retrospective ranks token-heavy hotspots using exact token accounting if it is visible in context. When telemetry is unavailable, it uses qualitative proxies like large file reads, long command output, broad searches, and repeated diffs, and labels them high, medium, or low.

No, it only works when the previous task is in the visible context. If the history is gone, you must supply a transcript, summary, branch, PR, or artifact before the skill can reflect, and it will not invent missing history.

Do not use it for planning before work begins, since it is strictly a retrospective tool. It is also unsuitable when no completed task exists in context or when you need precise billing-level token metrics that the environment does not expose.

It explicitly forbids fabricating numbers. If exact timing or token telemetry is absent, it states that once in the evidence basis and falls back to relative rankings with qualitative evidence such as repeated loops, large reads, or long-running tools.