Skip to content
Back to blog
Best Practices 5 min read

The knowledge hierarchy: how your AI gets smarter over time


Most AI coding tools start from zero every session. Your carefully crafted context vanishes the moment the terminal closes. The next developer on the team learns nothing from the last.

JuhJuh solves this with a four-level knowledge hierarchy that compounds institutional knowledge across every execution.

The four levels

  1. Organization: Standards that apply everywhere: coding conventions, security policies, and architectural principles. Set once, inherited by every project.
  2. Project: Project-specific context: domain models, API patterns, file structure, and deployment requirements. Each project carries its own knowledge base.
  3. Team: Team-level preferences and patterns. Frontend teams get different guidance than backend teams, even within the same project.
  4. Developer: Individual developer preferences. Your coding style, preferred patterns, and personal conventions layer on top of everything else.

How cascading works

Knowledge flows downward. Organization-level rules apply to every project unless a project overrides them. Project rules apply to every team unless a team overrides them. The most specific level wins.

Think of it like CSS: general rules cascade down, specific rules override. A developer's instruction about variable naming overrides the org default, but only for that developer's executions.

12 knowledge categories

JuhJuh organizes knowledge into 12 categories: architecture, API patterns, database schemas, deployment, testing, security, domain concepts, coding standards, dependencies, environment, performance, and custom. Each category can hold multiple entries, each with a scope and priority.

Knowledge extraction

When you connect a code repository, JuhJuh can extract knowledge automatically. It reads your project structure, identifies patterns, and populates the knowledge base. Subsequent executions use this knowledge to generate better prompts and more accurate code.

The result: every ticket your team completes makes the next ticket smarter. That is the juju.

See this in action.

The features described in this post are live in JuhJuh. Get started and explore the pipeline yourself.

Get started

Share this post
Post Share

From idea to pull request.

From reading to building. Get started on your first orchestrated pipeline.

JuhJuh uses cookies

By clicking “Accept all”, you agree to the storing of cookies on your device for functional, analytics, and marketing purposes. Only essential cookies are turned on by default. Cookie Policy

Essential

Required for the site to function

Analytics

Help us improve the product

Marketing

Personalized content and ads