pensieve

[!TIP]

Don't want to read docs? After installation, tell Claude: use loop to finish this task.

For Users Looking for the Linus Prompt

That methodology is still here and now delivered as an executable system:

• Linus-style engineering principles are embedded as default execution rules
• Core capabilities are mounted on the Claude Code runtime chain (skills / hooks / task / agent)
• Code review is shipped as review pipeline + knowledge

You get a full engineering package: prompting, workflow, and execution mechanics together.

What You Get Immediately

• Stronger breakthrough ability: stubborn bugs move toward root-cause fixes
• Higher code quality: cleaner, verifiable, maintainable output
• Harder code review: the review pipeline enforces a consistent standard
• More stable long sessions: on-demand loading keeps context lean
• Less manual babysitting: /loop splits, runs, and auto-continues tasks

30-Second Start

1. Install the plugin
2. Initialize .claude/pensieve/
3. Restart Claude Code
4. Type loop

Quick path (automation-first):

Share this README link or full text with your model and give this instruction:

Complete Pensieve installation and initialization based on this README, then report the result.

The model can execute installation, initialization, and verification steps automatically.

• Installation Guide
• Update Guide
• Upgrade Tool
• Uninstall Section

Why This Stays Reliable

1. Default principles drive execution

The system defaults prioritize root-cause repair, branch simplification, complexity control, and quality thresholds.

2. Context is loaded on demand

Each phase loads only what it needs: maxims for execution, review pipeline/knowledge for auditing, upgrade tool for migration.

3. Built on Claude Code native capabilities

• Skills: route intent to the right tool
• Hooks: SessionStart injects routes, Stop handles auto-continue
• Task: task state drives execution rhythm
• Agent: main window decomposes work, sub-agents execute

This design gives two direct benefits:

• Lighter architecture: native execution chain reduces wrapper complexity and maintenance overhead
• Upgrade compounding: Pensieve gains from Claude Code native improvements over time

Minimal Commands to Learn

• /loop: split complex work and execute in auto-loop mode
• /pipeline: list and invoke project pipelines
• /selfimprove: turn lessons into system-level improvements
• /upgrade: migrate and validate legacy structure to current standard

Best Fit Scenarios

• Stubborn bugs that keep recurring
• High-standard code review workflows
• Long-running, complex project iterations
• Teams that want to encode personal craft into default process

What Belongs in the Pensieve?

Maxim: should pass all checks. Still valid across projects? across languages? across domains? useful for future unknown problems?

Decision: keep it if any one of these is true. Would deleting it cause repeat mistakes? Would it help make a better choice in 3 months? Could you teach it as a reusable pattern?

Pipeline: any task shape that appears repeatedly. Make it runnable first, then refine.

How Memory Evolves

Temporary decisions in Loop -> filtered -> Decision
Repeated similar Decisions -> distilled -> Maxim
External knowledge + project practice -> Decision
Decisions guide -> Pipeline improvement

Like memory revealing truth, what you store helps Claude understand your intent.

Community

License

MIT