{"id":"fd1d8201-4f5b-4339-89b8-3a8734a3035a","slug":"retro-bot","kind":"skill","name":"retro-bot","summary":"Leave things better than you found them! This claude skill seamlessly integrates iterative feedback across your claude stack and keeps an audit log/paper trail.","owner":"cometogather","repoUrl":"https://github.com/cometogather/retro-bot","license":"MIT","licenseTier":"permissive","trust":"community","stars":14,"match":null,"fewStars":false,"format":"skill","createdAt":null,"addedAt":"2026-10-14T06:28:06.674Z","logoUrl":"https://avatars.githubusercontent.com/u/279766245?v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["developer-tools"],"tags":["agile","ai","claude","claude-code","retrospective","tokenmaxxing","continuous-improvement"],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-14T07:13:10.056Z","sample":false,"url":"https://conduits.dev/tools/retro-bot","language":"","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"```\n   .------------------------------------------------.\n   | retro-bot says: \"Always leave things better    |\n   | than you found them!\"                          |\n   '------------------------------------------------'\n        \\\n         |\n        [■]\n    .----------.\n    |[══]  [══]|\n    |          |\n    | .------. |\n    | |======| |\n    '----------'\n  .-------------.\n  |             |\n  | ✓  | ✗ | ↻ |\n  |   retro-bot |\n  '-------------'\n      ||   ||\n     _||   ||_\n    [__]   [__]\n```\n\n# retro-bot\n\n> *\"Always leave things better than you found them.\"*\n\n**An agile retrospective skill for your Claude collaboration sessions.**\nCapture what you learned. Apply the fix. Make the next session better.\n\n*Continuous improvement = continuous tokenmaxxing.*\n\n[![Install plugin](https://img.shields.io/badge/Install%20plugin-retro--bot%20v1.3-blue)](#install)\n[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](./LICENSE)\n[![Built with Claude](https://img.shields.io/badge/Built%20with-Claude-orange)](https://claude.ai)\n[![Ko-fi](https://img.shields.io/badge/Ko--fi-buy%20me%20a%20coffee-ff5e5b?logo=ko-fi&logoColor=white)](https://ko-fi.com/cj48744)\n\n<p align=\"center\">\n  <a href=\"https://ko-fi.com/cj48744\"><img src=\"https://ko-fi.com/img/githubbutton_sm.svg\" alt=\"Support me on Ko-fi\"></a>\n</p>\n\n---\n\n## How it fits into your workflow\n\nYou finish a Claude session — a document build, a data analysis, a code review, a content\ndraft, whatever the work was. Instead of closing the window and letting the friction\nevaporate, you type:\n\n> *\"let's do a retro on that session\"*\n\nTen minutes later: the fix is applied, the snapshot is saved, and the next session starts\nsmarter. That's it. retro-bot handles the rest.\n\n---\n\n## The problem\n\nIf you've worked in Agile, working with Claude should feel familiar. Especially as a\nproduct owner, your core skill is articulating what you want built, and that maps\ndirectly to getting great outputs from Claude. The clearer the brief, the better the\noutput. Same principle, different medium.\n\nI've found one valuable element from Agile that was missing for me: the retrospective.\n\nRetrospectives are the point of origination for iteration. It's where you clear the air,\nget on the same page, align on what worked and what didn't, and commit to the specific\nchanges that carry those lessons forward. It invites self-reflection and a collective\napproach to improvement. In a Claude workflow, its value can compound due to the more\nrapid cycles of creation and iteration.\n\nWithout a deliberate improvement loop, every session carries the same overhead. You\nre-establish context that should already be there. You hit the same friction you hit\nlast week. Learnings either evaporate or get dumped into one bloated file that creates\nits own problems. Token budgets inflate. Output quality plateaus, and you end up with\nan inflated, unwieldy prompt instead of a systemic approach to incremental improvement\nin the right places.\n\nMost people optimize the prompt. The best teams optimize the process.\n\nretro-bot brings the retrospective loop to Claude.\n\n---\n\n## What it does\n\nAfter any Claude session, retro-bot runs a structured retrospective with you, and\nthen does something most people haven't thought about: it routes each improvement to\nexactly the right layer of your Claude setup.\n\nThere are four places where context and instructions live in a Claude workflow:\n\n| Layer | What lives here |\n|---|---|\n| 🔧 Skill | How a specific capability is configured |\n| 📋 Project | What Claude knows about this engagement |\n| 📝 Claude.md | Persistent context that applies across all sessions |\n| 💡 Process | Behavioral agreements: how you and Claude work together |\n\nMost people know they should \"update their prompts\" after a session. What's less\nobvious is *where* the fix belongs. Without a clear answer, everything ends up\ndumped into one bloated global file. retro-bot figures out the right scope and applies\nthe change on the spot, versioned and recorded. That's the part most retro frameworks\nskip. retro-bot doesn't.\n\n```\n✓  What worked well      →  keep doing this\n✗  What didn't work      →  find the root cause\n↻  What to change        →  apply it to the right layer, right now\n```\n\nEvery retro saves a dated snapshot. A running audit log tracks every session, every\nimprovement, every outcome: a permanent record of how your Claude setup has evolved\nand why.\n\nAfter ten retros, you can look back and see exactly how your setup has evolved. The\nprompts are tighter. The skills are smarter. The sessions are faster. Your tokens go\nfurther. That's compounding. Call it tokenmaxxing.\n\n---\n\n## The audit trail\n\nThis is what your retro log looks like after a few sessions:\n\n```markdown\n# Retro Log\n\n| Date       | Session                         | Applied | Noted | Backlogged | Snapshot |\n|------------|---------------------------------|---------|-------|------------|----------|\n| 2026-05-15 | Doc build – long-form proposal  | 2       | 1     | 0          | [→](retros/2026-05-15-proposal-doc-retro.md) |\n| 2026-05-18 | Analysis – weekly metrics       | 2       | 0     | 1          | [→](retros/2026-05-18-metrics-analysis-retro.md) |\n| 2026-05-22 | Content draft – product brief   | 1       | 2     | 0          | [→](retros/2026-05-22-product-brief-retro.md) |\n| 2026-05-26 | Research – competitive review   | 3       | 0     | 1          | [→](retros/2026-05-26-competitive-research-retro.md) |\n```\n\nEach row links to a full snapshot: the retro board, every action item, what was applied,\nwhat was deferred, and any process agreements made. It's a record of how the workflow\nactually evolved: what stuck, what got reversed, what compounded. Nothing disappears.\n\n---\n\n## How a session looks\n\nYou finish a piece of work. You type:\n\n> *\"let's do a retro on that session\"*\n\nretro-bot goes first. Before asking for your input, it opens with blunt\nself-observations about what it noticed in the session. That's intentional: it breaks\nthe sycophancy default, models honest self-assessment, and usually surfaces things you\nnoticed but didn't say out loud. Then you build the board together, map issues to\nspecific file changes, apply them with versioning, and save the snapshot.\n\nA typical retro takes 10-15 minutes. The output is a versioned skill update, a Claude.md\naddition, or a process note, plus the dated snapshot in your archive.\n\n**Example exchange:**\n\n> **Claude:** Before you weigh in: here's what I noticed. The SOW table didn't render\n> cleanly in Word on the first pass; I had to rebuild it. That's a skill config issue I\n> can fix. I also wasn't sure which folder to save to until you told me mid-session.\n> A quick upfront question would have saved a round trip. On the positive side, the\n> staffing numbers stayed consistent across the scope spreadsheet and the SOW doc.\n> What's on your list?\n>\n> **You:** Yeah the folder thing kept happening last week too. Can we make that automatic?\n>\n> **Claude:** Good catch. That's a Claude.md update, not a skill update. I'll add your\n> default output path so it's set for every session. Proposing two changes total: patch\n> the table format in the SOW skill, and add the path preference to Claude.md. Want me\n> to apply both now?\n\n---\n\n## Retro formats\n\nThe default is WWW (What Worked / What Didn't / What to Change). Fast and practical\nfor most sessions. retro-bot also supports:\n\n- **Start / Stop / Continue**: when you want to explicitly introduce new behaviors\n- **4Ls: Liked, Learned, Lacked, Longed For**: when the session had a big learning component\n- **Mad / Sad / Glad**: when you need to name the frustration before solving it\n- **Sailboat**: for multi-session or quarterly reviews; strategic rather than tactical\n\n---\n\n## Prerequisites\n\nretro-bot is self-contained, with no dependencies on other skills or external services.\n\n**Required:**\n- **Claude Code** or **Cowork** (Claude desktop): the environments retro-bot runs in\n- **File system access**: so Claude can read and write skill files, instruction files, and retro snapshots\n  - *Cowork:* connect a folder via the workspace selector before running a retro\n  - *Claude Code:* file access is automatic within your project directory\n\n**Optional but recommended:**\n- [**Desktop Commander**](https://desktopcommander.app/): gives Claude broader file system access across your machine, useful if your skills or retro archive live outside the connected folder\n\nNo accounts. No APIs. No additional installs.\n\n---\n\n## Install\n\n### Cowork (Claude desktop)\n\n1. In Cowork, open **Customize → Plugins → +** → **Add marketplace**\n2. Paste this URL:\n\n   ```\n   https://github.com/cometogather/retro-bot\n   ```\n\n3. **Sync** the marketplace, then click **Install** on retro-bot in the catalog\n4. **Start a new chat** (or quit + relaunch Cowork). Skills don't hot-load into\n   an already-running session — if you try retro-bot in the same chat where you\n   installed it, Claude will say it can't find the skill.\n5. Done. Trigger it by saying *\"let's do a retro\"* after any session.\n\n**First run:** retro-bot will ask you to set an archive directory once, a folder\nwhere all retro snapshots will be saved automatically going forward.\n\n**Alternative (file install):** if you'd rather not add a marketplace, download\n[`retro-bot.plugin`](https://raw.githubusercontent.com/cometogather/retro-bot/main/retro-bot.plugin)\ndirectly and use Cowork's **Install plugin from file** option.\n\n### Claude Code\n\nInside an interactive Claude Code session:\n\n```\n/plugin marketplace add cometogather/retro-bot\n/plugin install retro-bot@retro-bot-marketplace\n```\n\nThe first command registers the retro-bot marketplace (one-time setup per machine).\nThe second installs the plugin. Trigger it by saying *\"let's do a retro\"* after any\nsession.\n\nTo update later: `/plugin marketplace update` then `/plugin upgrade retro-bot`.\n\n> **Heads up — HTTPS clone.** Claude Code defaults to SSH when cloning plugins from\n> GitHub. If you don't have a GitHub SSH key configured on your machine (most people\n> don't), install will fail with `Host key verification failed`. Set this environment\n> variable once before installing and you're good:\n>\n> ```bash\n> # macOS / Linux (add to ~/.zshrc or ~/.bashrc to persist)\n> export CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1\n> ```\n>\n> ```powershell\n> # Windows PowerShell (persistent, survives reboots)\n> [Environment]::SetEnvironmentVariable(\"CLAUDE_CODE_PLUGIN_PREFER_HTTPS\", \"1\", \"User\")\n> ```\n>\n> Then restart Claude Code and run the install commands above.\n\n### Claude Code (local development)\n\nIf you want to clone and run a local copy (for hacking on the plugin):\n\n```bash\ngit clone https://github.com/cometogather/retro-bot.git\nclaude --plugin-dir ./retro-bot\n```\n\n`--plugin-dir` loads the local directory as a plugin for the session without installing.\nUseful when you're modifying `SKILL.md` and want to test changes immediately. Run\n`/reload-plugins` inside the session after edits.\n\n### Having trouble installing?\n\nIf the marketplace install returns an error about source type compatibility, two fallbacks\nboth work on any Claude Code version with plugin support:\n\n```bash\n# Option 1: load locally for the session (no install)\ngit clone https://github.com/cometogather/retro-bot.git\nclaude --plugin-dir ./retro-bot\n\n# Option 2: download the plugin file and install from file\ncurl -LO https://raw.githubusercontent.com/cometogather/retro-bot/main/retro-bot.plugin\nclaude plugin install ./retro-bot.plugin\n```\n\nEither path bypasses the marketplace entirely.\n\n---\n\n## What's inside\n\n```\nretro-bot/\n├── .claude-plugin/\n│   ├── plugin.json                 ← plugin manifest (name, version, author)\n│   └── marketplace.json            ← marketplace listing so users can /plugin install\n├── skills/\n│   └── retro-bot/\n│       ├── SKILL.md                ← the skill (7-phase retro workflow)\n│       └── references/\n│           ├── retro-formats.md    ← WWW, 4Ls, Start/Stop/Continue, etc.\n│           └── action-item-types.md ← when to use each type of change\n├── .gitignore                      ← keeps personal retros and campaign assets out of the repo\n├── LICENSE                         ← MIT\n└── README.md                       ← you are here\n```\n\n---\n\n## The philosophy\n\nShared understanding breeds collective iteration.\n\nI built retro-bot because I noticed many components of Agile applied to how I\napproached my work with Claude. I found myself missing the retrospective loop I would\nrun in my IRL agile teams. I needed a structured way to make things better, get rid of\nrepetitive corrections, and continue the feedback loop.\n\nWhat wasn't obvious at first was what to do with the insights once you have them.\nClaude has multiple layers where improvements can live: the skill level, the project\nlevel, the global memory, the persistent context file. Getting the right fix to the\nright layer is what makes the improvement stick. That's the part most retro frameworks\nskip. retro-bot doesn't.\n\nI work with Claude every day at [Gather](https://gather.co): scope of work builds,\nbriefs, data analyses, content, research. The retro loop is how I make sure each\nsession teaches me something that makes the next one faster, sharper, and cheaper\nin tokens.\n\nI'm sharing this publicly because if it improves your outcomes, your token efficiency,\nand the quality of your outputs, we all stand to benefit. Shared understanding breeds\ncollective iteration. That only works if the loop is open.\n\nThis is a work in progress. Use it, fork it, improve it, run it back through itself,\nand pay it forward when you can.\n\n---\n\n## Version history\n\n| Version | Date | What changed |\n|---------|------|--------------|\n| v1.3 | 2026-05-29 | Migrated to plugin structure: added plugin.json, moved skill to skills/retro-bot/, repackaged as .plugin |\n| v1.2 | 2026-05-26 | Added project/session update type (project CLAUDE.md, client briefs, session configs) |\n| v1.1 | 2026-05-26 | Archive directory setup, mandatory snapshots, running audit log (retro-log.md) |\n| v1.0 | 2026-05-26 | Initial release: 7-phase retro workflow, 5 formats, copy-first versioning |\n\n---\n\n## Contributing\n\nFeedback is always welcome. If you use retro-bot and find something worth improving,\nopen a PR or an issue. Describe the friction, propose the fix. That's the whole\nmethodology applied to itself.\n\n---\n\n## If it made the next session better\n\nThis is free, MIT, and yours to fork. The premise is just that the friction you noticed\nat the end of a session is worth more than the session itself, and almost nobody writes\nit down. If retro-bot caught something you would have repeated, and you feel like saying\nthanks, there is a Ko-fi link below. Tips go into building more of these.\n\nIf it caught nothing, you owe me nothing. Seems fair.\n\n<p align=\"center\">\n  <a href=\"https://ko-fi.com/cj48744\"><img src=\"https://ko-fi.com/img/githubbutton_sm.svg\" alt=\"Support me on Ko-fi\"></a>\n</p>\n\nMore tools at [github.com/CR8-OS](https://github.com/CR8-OS).\n\n---\n\n## Built by\n\n**Christophe Jammet**, AI enablement and systems design  \n[linkedin.com/in/cjammet](https://linkedin.com/in/cjammet) · [github.com/CR8-OS](https://github.com/CR8-OS)\n\nPurpose built for [**Gather**](https://gather.co), a collective of independent thinkers,\ncreators and practitioners who move at the velocity your business demands.  \n[github.com/cometogather](https://github.com/cometogather)\n\n---\n\n## Work with Gather\n\n<a href=\"https://cometogather.kit.com/retro-bot\">\n  <img src=\"assets/lets-talk-card.png\" alt=\"Interested in bringing human-centered AI design to your organization? Let's talk.\" />\n</a>\n\n*retro-bot is just one example of human-centered AI in action. Want to see how we can bring it to your organization at scale?* **[Let's talk.](https://cometogather.kit.com/retro-bot)**\n\n---\n"}