Menu
Run end-of-session retrospective: compound learnings, then discuss skill performance, communication, and improvements.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Capture session learnings and save to skills, guidelines, or reference docs under ~/.claude/.
Orchestrate parallel claude -p sessions — bootstrap, launch, monitor, and converge. Works with any skill that produces manifest.json, item directories, and a runner script.
Create a request (pull request or merge request) or update an existing one following project conventions.
Resolve merge or rebase conflicts between branches.
Assess open PRs with unaddressed review comments and generate a parallel addressing script — produces manifest.json and let-it-rip.sh for address-request-comments execution.
Assess open work items and generate a parallel execution script — produces manifest.json and let-it-rip.sh for implement/clarify execution.
| name | session-retro |
| description | Run end-of-session retrospective: compound learnings, then discuss skill performance, communication, and improvements. |
| allowed-tools | ["Skill","AskUserQuestion"] |
Keywords: retrospective, end of session, learnings load review, provenance tracking, skill performance, communication quality, improvements, compound learnings, soft gate audit
End-of-session routine that captures learnings and facilitates a retrospective.
/session-retro - Run the full end-of-session retroCompound learnings: Invoke /learnings:compound using the Skill tool. Wait for it to complete fully before continuing.
Skill performance: Share your observations on how the skills used this session performed — what went well, what was rough. Include a learnings load review with provenance tracking — for each learnings file that was loaded or referenced, note how it was loaded:
| Source | Meaning |
|---|---|
| Index | Read via ~/.claude/learnings/CLAUDE.md index lookup |
| Hard gate | Session-start glob, plan-mode entry, or implementation-start check |
| Soft gate | Confidence-level check, friction-triggered, or keyword-triggered |
| Persona proactive | Listed in the active persona's ## Proactive Cross-Refs |
| Skill reference | Read because a skill's instructions or reference files pointed to it |
| Operator-prompted | Read because the operator asked a question or raised a topic that required it |
| Self-directed | Read on own initiative to inform a decision, without a gate or prompt triggering it |
| Cross-ref | Followed from a ## Cross-Refs in another loaded file |
Include a Loaded? column with three states to track actual context cost:
| Value | Meaning | Cost |
|---|---|---|
| Yes | Full read into context (by me directly, not an agent) | Real token cost |
| Sniffed | I directly issued a Read(file, limit=3) to check the header — not "an agent read it and I saw the summary" | Small cost |
| No | Known reference, never opened. Includes files read by subagents — their reads don't count as my context cost | Zero cost |
This distinction calibrates whether proactive cross-refs are earning their context cost.
This surfaces whether the search protocol is doing its job or whether useful files are only reached via persona coverage or manual reads. Also note: which files influenced decisions, which were noise, whether any useful files were missed, and whether any soft gates should have fired but didn't.
Then ask the operator for their perspective. Discuss naturally.
Communication quality: Share your honest assessment of how communication went — specific moments that worked, things you could've done better. Then ask the operator for their take. Discuss naturally.
Improvements: Share your ideas for improvements based on the above. As part of this, consider: did anything this session reveal a guideline that's missing, redundant, or out of sync with how we actually work? Then ask the operator if they have additional ideas. Discuss and capture anything actionable.
Final compound: Once all discussion is complete, invoke /learnings:compound again using the Skill tool. The retro discussion itself often produces new insights worth capturing — patterns noticed, guidelines proposed, communication preferences clarified. This second pass ensures those don't get lost.
Run each step sequentially. Lead with your own perspective at each stage, then invite the operator's — this is a two-way retro, not an interview. Do not advance to the next step until the operator signals they're done with the current one. Ask explicitly before moving on.
This is a conversation, not a facilitated meeting. The numbered steps are topics, not a script. Don't rush through them. Don't end every message with a formulaic handoff question ("What's your take?", "Ready to move on?"). Respond to what the operator actually said — follow up, push back, go deeper. Match the operator's pace and energy. If a topic has legs, stay on it. The steps exist so nothing gets skipped, not so you can optimize for completion.
Be verbose and genuinely reflective. Don't summarize — share your full thinking. Name specific moments, what worked, what you'd do differently, and why. The best retro insights come from detailed honest reflection, not polished bullet points. If you're holding back a thought, say it.