v3.6.0Open source · MIT
Your coding agent, with a method.
Playbook is a set of free plugins for Claude Code and Codex CLI. You decide what the software should do; the agent records it in your repo, builds from it, shows how it checked the result, and turns repeated mistakes into rules.
Works with Claude Code and Codex CLI
docs/wiki/works/playbook-product/decisions.md
- ID
- playbook-product-26
- Decision
- English at /, Spanish at /es/, Portuguese at /pt/, with automatic detection.
- Who decided
- owner
- Their words
- "si, ingles raiz, pero que detecte automaticamente"
- Kind
- definition
01The problem
Agents forget. Your repository should not.
Decisions vanish in the chat
What you agreed with the agent lives in a session that ends. Next week nobody knows why the code is the way it is.
Every session starts from zero
A teammate, or the agent itself tomorrow, cannot see what was decided, what changed, or which question is still open.
The same mistakes come back
Reviews catch the same kind of bug again and again, because nothing turns a lesson into a rule the agent follows.
02How it works
Six stages before any code. You close each one.
Every new feature or phase is a work with its own folder in docs/wiki/. The agent proposes, asks one question at a time, and records your answer word for word.
- 01
Wiki
What the product must do and must not do, who uses it, and every open question.
- 02
Technology
The stack, the limits, and what each choice costs, checked against current docs.
- 03
Skills
Which agent skills the work needs, and which ones exist already.
- 04
Design system
Tokens, components, and patterns, before a single screen is built.
- 05
Plan
Units of work, each with the paths it touches and how it will be accepted.
- 06
Execution
The agent builds unit by unit, records the evidence, and asks you to accept.
03It learns
Repeated mistakes become rules.
A review agent reads every risky change from a fresh context. Each defect it finds is stored as evidence under a kind of mistake. When three reviews hit the same kind, Playbook proposes one rule, and the agents that write and review code get it every time they start.
- 01review finds a defect
- 02classify stores it as evidence
- 03three hits on one kind
- 04a rule every agent follows
04Plugins
Three plugins. Take the wiki alone, or add the team.
playbook
The setup wizard, /playbook:init: choose the components and their scope.
Read the guide →playbook-wiki
Project memory: decisions kept in Git, each work prepared in six stages, checks on the wiki before commits and merges.
Read the guide →playbook-agents
A working team: explore maps code, write builds, review reviews, classify stores findings as evidence. Plus working rules and skills.
Read the guide →
05Install
Paste one prompt. Your agent does the rest.
Copy the prompt for your client into a session. Then run /playbook:init in your project.
Claude Code
Install Playbook for Claude Code from https://github.com/andresfrei/playbook. Add its marketplace with `claude plugin marketplace add andresfrei/playbook` if it is missing. Run `claude plugin install playbook@playbook --scope user`. Use `claude plugin list` to confirm the plugin is enabled and tell me its version. Then tell me how to start a fresh Claude Code session in my project and run /playbook:init so I can choose the components and scope.
Codex CLI
Install Playbook for Codex CLI from https://github.com/andresfrei/playbook. Ask me which setup I want: 1 for the wiki only, or 2 for the wiki and agents. Use my local Playbook checkout if it contains codex/install.py. Otherwise, clone the repository into a fresh directory. From the checkout root, run `python3 codex/install.py` and supply my choice when it prompts. Report what was installed and any existing profiles the installer kept. Show me which hooks to review in /hooks, then how to start a fresh Codex session in my project.
06News
It grows by building itself.
Every release since the first, read from the repository. Each review finding Playbook records can become a rule it ships in the next version.
- releases
- rules learned
19 releases13 rules learned
See the data as a table
| Version | Date | releases | rules learned |
|---|---|---|---|
| 1.0.0 | Nov 14, 2026 | 1 | 0 |
| 2.0.0 | Nov 17, 2026 | 2 | 0 |
| 2.1.0 | Nov 17, 2026 | 3 | 0 |
| 2.2.0 | Nov 17, 2026 | 4 | 0 |
| 2.2.1 | Nov 17, 2026 | 5 | 0 |
| 2.2.2 | Nov 17, 2026 | 6 | 0 |
| 2.3.0 | Nov 18, 2026 | 7 | 0 |
| 2.4.0 | Nov 18, 2026 | 8 | 0 |
| 2.5.0 | Nov 19, 2026 | 9 | 0 |
| 2.6.0 | Nov 19, 2026 | 10 | 0 |
| 3.0.0 | Nov 19, 2026 | 11 | 0 |
| 3.1.0 | Nov 19, 2026 | 12 | 0 |
| 3.1.1 | Nov 19, 2026 | 13 | 0 |
| 3.2.0 | Nov 19, 2026 | 14 | 0 |
| 3.2.1 | Nov 19, 2026 | 15 | 0 |
| 3.3.0 | Nov 20, 2026 | 16 | 0 |
| 3.4.0 | Nov 20, 2026 | 17 | 4 |
| 3.5.0 | Nov 20, 2026 | 18 | 9 |
| 3.6.0 | Nov 20, 2026 | 19 | 13 |
Release notes in English, as published on GitHub.
v3.6.0Nov 20, 2026
Release notes →Dedicated agents for repeated tasks
At the close of each work, Playbook now records the tasks the agent carried out, and proposes a dedicated agent, on a model that fits the task, for a task that repeats (work task-agents, #44).
v3.5.0Nov 20, 2026
Release notes →Units built in parallel
A work's plan now says which units can be built at the same time, and stage 6 builds them in parallel (work parallel-units, #42).
v3.4.0Nov 20, 2026
Release notes →Ask before what cannot be undone
Claude Code now asks you before the actions you cannot take back, and a new command protects a repository's secrets (work ask-before-irreversible, #40).
See all 19 releases ↓Show fewer ↑
v3.3.0Nov 20, 2026
Release notes →Evidence seeded from past reviews
A repository can now start its evidence page from the reviews it already had (work seed-evidence, #38).
v3.2.1Nov 19, 2026
Release notes →playbook-agents: when a review ends with findings and starts no classify, the hook now stops it once per subagent.
Before, stop_hook_active was not set on the repeated stops of one review, so a review that could not start the classifier was stopped again and again (repeated-corrections-31, #36).
v3.2.0Nov 19, 2026
Release notes →Learned rules from repeated review findings
When review keeps finding the same kind of mistake, the agent now learns from it (work repeated-corrections, #34).
v3.1.1Nov 19, 2026
Release notes →A fix to the version notice of 3.1.0.
v3.1.0Nov 19, 2026
Release notes →Claude Code does not tell you when a plugin from this marketplace has a new version.
From 3.1.0, the playbook tells you itself, at the start of a session, and the notice about your wiki's format now reaches you, not only the agent.
v3.0.0Nov 19, 2026
Release notes →Version 3 lets you delegate a kind of question, so the agent stops asking the ones you always answer "yes" to.
It is a major version: it moves a project's wiki to format 5.
v2.6.0Nov 19, 2026
Release notes →prepare-project now asks how a work is delivered, before any code.
v2.5.0Nov 19, 2026
Release notes →Three commands to start and configure Playbook in Claude Code.
v2.4.0Nov 18, 2026
Release notes →Playbook now supports Codex CLI alongside Claude Code.
v2.3.0Nov 18, 2026
Release notes →Version 2.3.0 makes the agent review read a risky change before it is committed, instead of relying on the agent to remember rule R20.
It adds two hooks to playbook-agents and breaks nothing.
v2.2.2Nov 17, 2026
Release notes →Version 2.2.2 makes the report of review-change and the questions of prepare-project closer to what each skill says, and makes their evaluation cases read the meaning of an answer.
It breaks nothing.
v2.2.1Nov 17, 2026
Release notes →Version 2.2.1 fixes skills that did not load when a request called for them, and a script that failed in a sandbox.
It breaks nothing.
v2.2.0Nov 17, 2026
Release notes →Version 2.2.0 brings the skills in line with the best practices for skills that Anthropic and agentskills.io publish, and breaks nothing.
v2.1.0Nov 17, 2026
Release notes →Version 2.1.0 adds two things to playbook-wiki and breaks nothing.
v2.0.0Nov 17, 2026
Release notes →Version 2 splits the playbook into a shared core, for everyone who works on a project, and the owner's agent system, which anyone may use.
It is a major version: it breaks an installation of version 1.
v1.0.0Nov 14, 2026
Release notes →First public release of the playbook: one person's working rules, skills, and agents for AI coding agents, shared as it is.
07Open source
Free, MIT, and built with itself.
Playbook is developed with Playbook: its own repository holds the decisions, the plans, and the evidence that shaped every release. Read them, fork it, open an issue.
- Current version
- v3.6.0
- License
- MIT
08Coming soon
A store of agents and skills.
Playbook already turns repeated mistakes into rules and creates dedicated agents for tasks that repeat. Next: a place to share them. It is not open yet.
- 01
Back up
Keep your global agents and skills safe, not only on your machine.
- 02
Publish
Share an agent or a skill that works, so others can install it.
- 03
Install
Browse what others shared and add it to your setup.