debrieflisted

# Session Retrospective You are reviewing THIS session (the conversation so far) — not any file, diagram, feature, or piece of code it produced. The goal is process improvement: how you (the agent) worked, not what got built. Do not evaluate the quality of any deliverable; that's a different kind of review. ## 0. Load context first Look back over the actual conversation in this session — the real sequence of tool calls, questions asked, decisions made, corrections or confirmations the user gave. Don't reconstruct this from assumption or general impression; look at what actually happened. If a CLAUDE.md, README, or project memory exists for the current repo, skim it so you know what standing conventions were already supposed to be followed — but don't assume any of these exist. **Scoping note:** for a very long session, reading the entire transcript isn't free. Focus on the most relevant or most recent stretch, or ask the user which part they want reviewed, rather than retro-ing the whole thing by default. ## 1. Efficiency and token usage - Which steps consumed unnecessary tokens — re-reading files that hadn't changed, redundant tool calls, guessing at something checkable instead of checking it, more cycles (build/render/test/ whatever the domain's loop is) than needed, restating information already established earlier in the conversation? - **Errored vs. successful repetition, distinct from the above:** an errored tool call is self-flagging — it produces visible