claude-bayesian-playbook

bayesian-it

A Bayesian learning engine that makes Claude Code progressively smarter in your codebase.

---

Every coding session is a data point. bayesian-it silently captures episodes from your Claude Code sessions, then uses counterfactual reasoning to extract patterns correlated with success and anti-patterns correlated with failure. These are stored in a Markdown playbook with Bayesian confidence scores that update as more evidence accumulates. High-confidence patterns are automatically injected into new sessions, so Claude's advice gets sharper the more you use it.

How It Works

  Claude Code Session
        |
        v
  +--------------+     hooks (PostToolUse / Stop)
  |   Episode    |<---- silently captures every tool call
  |   Capture    |      and session outcome
  +------+-------+
         |
         v
  +--------------+     on-demand via /bayesian:analyze
  |   Scribe     |<---- LLM-powered analysis with
  |   Analysis   |      counterfactual reasoning
  +------+-------+
         |
         v
  +--------------+     .bayesian/playbook/*.md
  |  Playbook    |<---- human-readable Markdown patterns
  |   Store      |      with YAML frontmatter + confidence
  +------+-------+
         |
         v
  +--------------+     auto-injected at session start
  |  Claude      |<---- high-confidence patterns surfaced
  |  Context     |      automatically + on-demand via commands
  +--------------+

Features

Feature	Description
Bayesian Confidence	Wilson score interval calculations with observation-count thresholds (low -> very_high)
Counterfactual Reasoning	The Scribe classifies causation as causal, likely_causal, correlational, or spurious -- spurious patterns are auto-filtered
Cold Start Mitigation	12 pre-seeded language-agnostic patterns across testing, debugging, refactoring, and error handling
Source Provenance	Every pattern tagged as pre-seeded or learned so you know what came from evidence
Silent Hooks	Episode capture never blocks Claude Code -- errors are swallowed, exit code is always 0
Markdown Playbook	Patterns stored as .md files with YAML frontmatter -- human-readable, git-diffable, versionable
Auto-Injection	High-confidence patterns automatically included in session context at startup
Slash Commands	5 built-in commands for interacting with the playbook without leaving your flow

Installation

Prerequisites

• Node.js >= 18
• Claude Code CLI
• ANTHROPIC_API_KEY environment variable (for the Scribe's LLM analysis calls)

Global install (recommended)

Installs hooks, MCP server, and commands into your global Claude Code configuration. Works across all projects.

npx claude-bayesian-playbook --global

Project-scoped install

Installs into the current project only.

npx claude-bayesian-playbook --local

Verify

Start a Claude Code session and run:

/bayesian:status

You should see episode counts, pattern stats, and confidence distribution.

Commands

Command	What it does
/bayesian:status	Dashboard -- episode count, pattern stats, confidence distribution, recent Scribe runs
/bayesian:query	Query the playbook for patterns relevant to a task or context type
/bayesian:analyze	Trigger Scribe analysis and see a before/after summary
/bayesian:explain	Deep-dive into a pattern -- Wilson score math, causation assessment, full evidence
/bayesian:review	Batch review workflow -- approve, reject, or skip patterns one by one

Uninstall

npx claude-bayesian-playbook --uninstall

Removes all bayesian-it entries from settings.json, .mcp.json, and CLAUDE.md without affecting other plugins. Episode data in .bayesian/ is preserved.

Confidence System

Confidence is calculated using the Wilson score interval lower bound:

lower = (p + z^2/2n - z*sqrt((p(1-p) + z^2/4n) / n)) / (1 + z^2/n)

Where z = 1.96 (95% CI), p = success_rate, n = total_observations.

Level	Score Threshold	Min Observations
very_high	>= 0.75	10
high	>= 0.55	5
medium	>= 0.35	3
low	below thresholds	--

For anti-patterns, the failure rate is used instead of the success rate.

License

MIT