Post-session skill turns the session on itself. What shape did the work take, where did the context go, what should the next session inherit. No flattery.
A skill that reviews how the current session was conducted — not what it produced — from the conversation already in context. It classifies the session against a fixed taxonomy of shapes, scores how well it was run, names where the context window was spent badly, and turns repeatable hand-work into concrete improvements to the skills and workflows it used — or, when the cost would recur for any agent, names the environment lever that removes it at the source.
This skill follows the Agent Skills specification so it can be used by any skills-compatible agent.
npx skills add robcsaszar/retrospective
/plugin marketplace add robcsaszar/retrospective
/plugin install robcsaszar-retrospective@retrospective
Copy the skills/retrospective/ directory into your project's .claude/skills/.
Run it at the end of a working session with /retrospective (optionally --focus spend or --focus wiring). It is user-invoked by design — a retrospective is a deliberate act, not something to fire on its own.
The premise that makes it cheap and honest: the session is already the skill's context. There is no log to parse and no gathering step, so the review is grounded in what actually happened rather than inferred from a tool-call tally — and it never opens an on-disk transcript, including one handed to it as "the session to review."
It does three things. It takes the session's shape, classifying it as exactly one of six — Scope-then-Build, Run-the-Plan, Survey-then-Propose, Pinpoint Fix, Orchestrated Run, Contested Inquiry — and scoring it 1–10 against a shared key, so the classification is consistent across runs instead of a fresh label each time. It accounts for the context, naming the concrete moments where tokens were spent badly, each tied to one of five named economy rules with the specific different action for next turn, reasoned in relative magnitude and never a fabricated count. And it wires the learning back, folding a repeatable step performed by hand into the skill or workflow the session actually called, or briefing a new skill when nothing existing fits — written to a durable handoff note, best-effort, so the next session inherits it.
Since 0.2.0 the wiring step asks one more question first: was it the turn or the ground? A cost a more careful turn would have avoided gets a next-turn fix. A cost the environment imposes on anyone — a file that is hard to find, a mistake no check catches, a tool that is verbose by design, a fact the agent cannot reach — gets one of six named levers (navigation pointer, automated check, review-stage rule, steering trim, tool verbosity, information access) and the real file it lands in, most severe first.
MIT © Rob Csaszar