session-retro
Skill1 starsChanged 15 days agoMITView on GitHub
Session insights report. Finds what cost this session time, technical or not: decisions that reversed, terminology churn, misread intent, corrections the user had to repeat, and why the handoff wasn't taken when ccBar offered one. Writes the report to ~/.claude/retros/ and proposes durable lessons as memory or CLAUDE.md edits for approval. Triggered automatically by ccBar's Stop hook after the handoff threshold, or run manually any time.
15 other projects use a version of this.
Adopted by
Most recently taken up by
- jatinmishra/claude-skills-monorepotheir own version, first seen Nov 8, 2026
- phntm7/skillstheir own version, first seen Nov 1, 2026
- momentmaker/kaijutsutheir own version, first seen Nov 1, 2026
- adeonir/agent-skillstheir own version, first seen Nov 1, 2026
- davisbuilds/dojotheir own version, first seen Nov 1, 2026
- TheAxeC/ordotheir own version, first seen Nov 1, 2026
Other versions15 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 15
- werf/werf★ 4.7ksame name
- Arman-Kudaibergenov/1c-ai-development-kit★ 163same name
- adeonir/agent-skills★ 8same name
- renzrollon/shippable-skills★ 7same name
- Zejnilovic/agd★ 5same name
- momentmaker/kaijutsu★ 3same name
- bbeny123/pragma★ 2same name
- Kwangseok-Seo/skills★ 2same name
- nhhthong/skills★ 1same name
- phntm7/skills★ 1same name
5 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository2 other files it gives its agents.
- ccbar-tuneSkill
- session-handoffSkill
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 4 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 14 edited versions we read keep these.
4
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 28
Only if the context shows a compaction summary, use Grep on transcript_path for specific evidence: tool errors, the user's corrections …
- Distinctive: line 42
Keep the three to five findings that cost the most.
- Very specific: line 46
Write it to ~/.claude/retros/YYYY-MM-DD-<cwd leaf>-<first 8 chars of session_id>.md, using today's date.
- Distinctive: line 86
Do NOT write any memory, CLAUDE.md or decisions-log change until the user approves it by number.
- Distinctive: line 90
After reporting, stop. Don't resume the prior task until the user says so.
- Distinctive: line 92
Write in the user's style, following any writing rules in their CLAUDE.md files or memory.
“Common” is measured: kept by at least 6 in 10 of the 14 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/session-retro
Blunt, evidence-based insights from this session so far. Args (when triggered by ccBar): session_id, transcript_path, cwd, and handoff_fired_at (tokens) when ccBar's handoff already fired this session.
1. Source
Review the conversation already in your context. Do NOT Read the transcript file: it is several MB at this size and would pull a large part of the session back into context.
Only if the context shows a compaction summary, use Grep on transcript_path for specific evidence: tool errors, the user's corrections ("no", "don't", "wrong", "stop", "I meant", "actually", "again"), renames, reverted edits. Never read it whole.
2. Find what cost time
Most of a session's lessons are not technical. They're about what words meant, what was decided, and how the user and Claude worked together. Scan with each lens:
- Skipped handoff (only when
handoff_fired_atis set): ccBar offered a handoff and the session kept going. Look at what happened after the handoff plan was written. Was the work mid-step, so a cut would have cost more than it saved? Did the user choose to continue? Did scope grow, or did the session drift into side work? Say plainly whether continuing was worth it and where a clean cut was. If the handoff plan was written in the immediately previous response, the user hasn't had a chance to act yet: say so in one line instead. - Decisions that moved: a product, design, naming or technical decision that was made, then reopened or reversed. Note what it finally settled on, how many rounds it took, and what question or fact, raised at the start, would have settled it in one.
- Language and intent: terms the user and Claude used differently, wording that churned (labels, copy, names), asks read too literally or too loosely, things the user had to say twice.
- Working style: asked when it should have acted, or acted when it should have asked; did more or less than asked; length or tone the user pushed back on.
- Technical misses: wrong assumptions about code, tools or environment; verification that failed or was skipped; tool errors that cost a retry.
Also note an approach the user explicitly confirmed when it wasn't the obvious choice. That's a lesson too. Routine success isn't.
Keep the three to five findings that cost the most. Don't pad to fill a lens, and don't invent problems.
3. Write the retro
Write it to ~/.claude/retros/YYYY-MM-DD-<cwd leaf>-<first 8 chars of session_id>.md, using today's date. On a manual run with no session_id, use HHMM in place of the session fragment.
# Retro: <what the session was for, one line>
## Headline
<One to three sentences: the biggest cost this session and its cause. If nothing notable, say so and stop here.>
## Skipped handoff
<Only when handoff_fired_at is set.>
## Findings
### <Lens>: <short name>
- **What happened:** <evidence: the user's words quoted briefly, or the file or tool call>
- **Cost:** <rough turns or time lost>
- **Lesson:** <what to do differently, as a rule someone could follow>
## Confirmed
<Only if any.>Order findings by cost, highest first. Quote the user verbatim and briefly; never paste code, secrets or long passages.
4. Durable lessons
For each lesson likely to recur, pick the single most appropriate destination:
- Auto-memory
projectentry: a product, design or terminology decision the session settled after back-and-forth, including what was rejected, so later sessions don't reopen it. If the session worked from a project decisions log (a PM handoff or product docs), propose a row there instead. - Auto-memory
feedbackentry: how the user wants Claude to work, personal and cross-session, not already in a CLAUDE.md. - Workspace or project
CLAUDE.md: the nearest existingCLAUDE.mdwalking up fromcwd, preferring one above the git repo root when it exists. Never propose creating a newCLAUDE.mdinside a git repo. Before proposing an edit to one inside a repo, check~/.claude/CLAUDE.mdand any ancestorCLAUDE.mdfor a rule against repo-levelCLAUDE.mdfiles; if there is one, use the next destination up. - Global
~/.claude/CLAUDE.md: a rule about how the user works across all projects.
Memory entries follow the memory file format from your system prompt (one fact per file plus a MEMORY.md pointer).
Before proposing, read the target and check for an existing entry that covers it. Propose updating that entry rather than adding a duplicate. If the lesson already exists as a rule, say the session didn't follow it instead of proposing a duplicate. One-off mistakes don't become rules. Expect 0 to 3 proposals.
5. Output in chat
- The headline
- The retro file path
- Numbered proposals, each with the target path and the exact text to add or change
Do NOT write any memory, CLAUDE.md or decisions-log change until the user approves it by number.
6. Stop
After reporting, stop. Don't resume the prior task until the user says so.
Write in the user's style, following any writing rules in their CLAUDE.md files or memory.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

