reflect

Invoke the reflect skill after a complex task completes. It locates the active transcript, spawns three parallel reviewer subagents with different lenses, synthesizes their findings, and proposes concrete edits to existing skills for your approval.

Each reviewer applies a distinct lens: judgment names durable principles behind incidents, tooling captures concrete commands and technical facts, and divergent finds blind spots and second-order effects. Their outputs are merged by a synthesizer subagent.

No. The synthesizer produces an Accepted, Rejected, and Backlog list, and the parent agent presents it for explicit user approval before applying any edit. Only approved items are applied, routed either as direct edits or through the create-skill workflow.

Skip reflection when the conversation was trivial, off-topic, or already covered by an existing skill the agent followed correctly. One-off events and implementation details that drift, like SHAs or version numbers, are not durable learnings.

Findings better enforced by a lint rule, script, metadata flag, or runtime check are moved to the Backlog list. The parent files those items to the team's devex or backlog tracker rather than editing skill prose.