Catches the wasted tokens, mistimed questions, and repeated mistakes a code review can't see. A single-file Claude Code skill for session retrospectives, with no scripts or dependencies.
Origin story (soon): We Review Our Code. Why Don't We Review How We Build It With AI?
- Use when wrapping up a Claude Code session, or when asked for a debrief, a session retrospective, or a post-mortem.
- Pairs with code review: code review evaluates the implementation;
/debriefevaluates how the session that produced it went.
/plugin marketplace add Wavez/debrief-skill
/plugin install debrief-skill@debrief-skill
Tip
To auto-update: run /plugin → Marketplaces → select debrief-skill → Enable
auto-update.
Run /debrief at the end of a session so Claude stops repeating the same process mistakes and
gets better at working with you over time.
Read-only and self-contained: it never edits files or saves memory unless you explicitly approve it.
- Efficiency and token usage
- Redundant tool calls, unnecessary re-reads, and missed batching or parallelization, so you catch wasted tokens and time before they repeat next session.
- Overlapping-dispatch: two subagents that independently redid each other's work.
- Context-seam: whether anything had to be re-established after a long session's context got summarized.
- Collaboration and communication quality
- Questions asked when needed and skipped when the answer was already checkable, standing instructions honored or missed, and whether replies stayed concise.
- Timing check: a question that's reasonable in content but lands before you've signaled you're done giving feedback.
- Correctness and process discipline
- Unverified claims, hard-to-reverse actions taken without confirming, overlooked project conventions.
- Constraint-timing: whether hard requirements were confirmed before finalizing a plan, or only discovered after the work was done.
- Follow-through for next time
- What's worth saving from this session so it doesn't need to be re-learned next time.
- Regression check: cross-references findings against saved feedback/memory, so a repeated mistake reads as a regression, not a first-time miss.
- Missing tools or scripts: hand-writing the same logic twice in a session is itself flagged as a signal that a reusable script or skill asset is missing.
Overall assessment Mostly efficient session, one timing slip worth flagging.
Concrete issues
🔴 High: Claimed the fix was "done" before running the test suite.
🟡 Medium: Asked whether the architecture was finalized while you were still listing constraints.
- A mistimed question — reasonable in content, asked before you'd signaled you were done giving feedback.
🟢 Low: Re-read config.py three times after no edits had been made to it.
What went well
- Batched the four independent file edits into a single parallel tool call.
💡 Suggested improvements
- Project memory — record the "claimed done before testing" incident (the High finding above), so a repeat of it next session reads as a regression, not a first-time miss.
- Global
CLAUDE.md— from the Medium finding above: wait for an explicit "done giving feedback" signal before asking the user to decide between options, even when the question itself is reasonable. General collaboration habit, not specific to this repo.
If you already use a different retrospective tool, here's how this one differs:
| Alternative | What it does differently |
|---|---|
bitwarden/ai-plugins claude-retrospective |
Heavier: pulls git history and examines existing code-quality metrics via companion shell scripts from Bitwarden's plugin ecosystem. A full session audit, not a lightweight process check. |
TerenceBristol/claude-improve |
A continuous, multi-run configuration-improvement system tracking acceptance rates across sessions over time — an ongoing machine, not a single end-of-session retrospective. |
session-report (official Anthropic plugin) |
Narrative call-outs on token waste and cost anomalies from usage data — not a judgment of the collaboration or process itself. |
See CHANGELOG.md for release history.