Skip to main content

​
Overview

Self-improving packages enable AI assistants (Claude, Cursor, etc.) to automatically search and install PRPM packages when they detect they’re working on specialized tasks. Instead of relying only on their base training, the AI can dynamically acquire domain-specific expertise from the PRPM registry.

​
How It Works

​
1. Task Detection

The AI analyzes the user’s request for domain-specific keywords:
  • Infrastructure: aws, pulumi, terraform, kubernetes, docker, beanstalk
  • Testing: playwright, jest, cypress, vitest, e2e
  • Deployment: ci/cd, github-actions, gitlab-ci, workflows
  • Frameworks: react, vue, next.js, express, django, fastapi
When keywords are detected, the AI searches the PRPM registry:

​
3. Quality Assessment

The AI evaluates packages based on confidence levels: High Confidence (Auto-suggest)
  • ✅ Official packages (@prpm/*)
  • ✅ Featured packages
  • ✅ High downloads (>1,000)
  • ✅ Verified authors
Medium Confidence (Present options)
  • ⚠️ Community packages (<1,000 downloads)
  • ⚠️ Multiple similar packages
Low Confidence (Skip)
  • ❌ Unverified packages
  • ❌ Zero downloads

​
4. User Approval

The AI presents findings and asks permission:

​
5. Installation & Application

After approval, the package is installed and immediately applied:
The AI then uses the package’s expertise to help with the task.

​
Real-World Example

​
User Request

​
AI Detection & Response

​
After Installation

The AI now has access to:
  • Production-tested Pulumi patterns
  • Beanstalk deployment best practices
  • Common pitfalls and solutions
  • CI/CD integration examples

​
Installing Self-Improving Packages

​
For Claude Code

This skill teaches Claude to:
  • Detect task-specific keywords
  • Search the registry proactively
  • Evaluate package quality
  • Request permission before installing
  • Apply package knowledge immediately

​
For Cursor

This rule enables Cursor to:
  • Automatically trigger searches on infrastructure/testing tasks
  • Present top packages with download counts
  • Install packages after user approval
  • Load expertise for the current task

​
Creating Your Own Self-Improving Package

You can create packages that teach AI assistants to be self-improving:
Publish it:

​
Meta-Dogfooding

PRPM uses its own self-improving packages for development:
  • @prpm/pulumi-infrastructure → PRPM’s infrastructure (74% cost savings)
  • @prpm/github-actions-testing → PRPM’s workflow validation
  • @prpm/postgres-migrations → PRPM’s database patterns
Benefit: Users get the same battle-tested expertise that built PRPM itself.

​
Privacy & Security

  • ✅ All searches query the PRPM registry API
  • ✅ No personal data collected during searches
  • ✅ Download tracking only on install (anonymous)
  • ✅ User approval required before installation
  • ✅ Packages are scanned for quality and safety

​
Best Practices

​
1. Be Proactive

Search for packages before starting complex tasks, not after encountering errors.

​
2. Verify Quality

Always check:
  • Download counts
  • Official/verified status
  • Package description relevance

​
3. Ask Permission

Never install packages without explicit user approval.

​
4. Apply Immediately

Once installed, use the package knowledge right away on the current task.

​
5. Track Helpfulness

Note which packages were useful for future reference.

​
Advanced Patterns

​
Chaining Package Discovery

AI can discover packages that help discover more packages:

​
Context-Aware Installation

AI can choose package format based on the editor:

​
Temporary vs Permanent

Some packages are task-specific, others are project-wide:

​
Limitations

  1. Search Quality: Results depend on package descriptions and tags
  2. Installation Friction: Requires user approval (by design)
  3. Context Window: Very large packages may exceed token limits
  4. Network Required: Registry search needs internet access