All directories
SkillProductGrowth

Self-Improving Knowledge System

File-based knowledge graph pattern (Product Compass) — CLAUDE.md brain, INDEX.md router, progressive disclosure domains, hypothesis tracking, and compounding agent workflows.

Open sourceBrowse more skills
01

Overview

Paweł Huryn’s self-improving Claude system pattern: pull data → organize knowledge → let the system learn → compound. Architecture is a file-based knowledge graph with progressive disclosure — CLAUDE.md as operating brain, knowledge/INDEX.md as router, domain folders (craft/voice/platforms/hypotheses or PM equivalents like discovery/stakeholders/channels). Works across Claude Code, Cowork, and web with shared context. Includes hypothesis tracking and “false beliefs” to keep the system honest. Adapt labels to customer research, competitive intel, or market analysis — not only content.

SaaSAgencyEducationOther

Work with LimeDock

Using this skill? LimeDock can wire it into a durable automation you own.

Skills show what's possible. LimeDock builds and runs the owned marketing, sales, and ops automations around them.

We sell owned automations for SaaS teams — live workflows that plug into Slack, CRM, and your internal platform — not just a skill list.

Book a workflow call
02

Link

https://www.productcompass.pm/p/self-improving-claude-system
03

Installation guide

There is no single npm package — scaffold the structure in your repo:

CLAUDE.md
knowledge/
  INDEX.md
  craft/          # or discovery/
  voice/          # or stakeholders/
  platforms/      # or channels/
  posts/          # performance data
  hypotheses/

Paste the Product Compass architecture into Claude and ask it to generate the starter files for your domain. Optionally add Python fetch scripts as the system proposes them.

04

How to use it

1. Start messy (screenshots, raw notes). 2. Let Claude propose hierarchy once patterns emerge. 3. Keep INDEX.md as the only always-loaded router. 4. Track hypotheses with kill criteria; log false beliefs. 5. Reuse the same files from Code, Cowork, and web.

05

Example prompts

  • “Scaffold a Product Compass-style knowledge system for B2B customer interviews.”
  • “Reorganize these notes into knowledge/ with INDEX.md progressive disclosure.”
  • “Add hypothesis tracking and false-beliefs files; propose 5 tests for our onboarding copy.”
  • “Summarize who decided what in this session and update CLAUDE.md learning rules.”
06

Use cases and examples

  • Compounding content / social system with real performance data
  • Customer-interview knowledge base that improves after every call
  • Competitive monitoring with weekly pattern synthesis
  • PM operating system across Claude surfaces
  • Replace cold-start chats with progressive disclosure
07

Prerequisites

  • Git repo where CLAUDE.md + knowledge/ can live
  • Claude Code and/or Cowork access
  • Willingness to own editorial/judgment calls — AI compresses execution, not taste
08

Tips

  • Don’t load the whole graph every turn — router first.
  • Cross-surface shared files beat per-tool prompt paste.
  • Pair with claude-mem for automatic capture; keep the graph for curated truth.
  • Related: Candlekeep / architecture explainers on Claude Code internals as inspiration, not required deps.
09

Sources