Using AI with AppCat

The AppCat repository includes configuration for AI-assisted development using Claude Code. Each AppCat service has its own dedicated agent configuration that acts as a domain-specific knowledge worker and coding buddy.

This approach has been tested with Claude Code and no other AI coding tools at this time.

How it works

The AppCat repo contains a CLAUDE.md and per-service agent configurations. Agents serve as reviewers and advisors for the main coding session:

  • Each agent is scoped to a specific AppCat service, keeping its context focused and relevant.

  • The main session can bounce ideas and implementation approaches off the agent.

  • This creates a continuous review loop without requiring the full project context in every session.

  • Agents guide the main session and act as knowledge workers, they don’t write code themselves.

Dos and don’ts

Do

  • Use AI to assist your work: brainstorming, drafting, reviewing, debugging.

  • Review all AI output before using it. Ensure it is correct and meets our standards.

  • Keep your own understanding of the code you’re working on. If you can’t explain it, you don’t own it.

Don’t

  • Don’t blindly accept AI output. You are responsible for the code you submit.

  • Don’t submit AI-written code for review without reviewing it yourself first.

  • Don’t blindly trust AI-generated tests. Verify they actually test meaningful behavior and not just the implementation.

Recommended plugins

The following Claude Code plugins are recommended to improve the AI-assisted workflow:

Caveman

Compresses Claude’s output by cutting filler words. Gets straight to the point and saves tokens.

Code Review

Structured code review with a scoring system that filters out false positives.

Superpowers

Adds structured workflows to Claude Code. Starts by clarifying intent, builds a spec, then creates an implementation plan emphasizing TDD and YAGNI. Launches sub-agents to work through tasks with built-in review checkpoints, enabling longer autonomous sessions without deviating from the plan.

Usecases

Usecases for AI with AppCat:

Code Reviews

Claude can assist in code reviews and help spot hard-to-find bugs. To filter out false positives, use the Code Review plugin described above. It can be invoked using /code-review:code-review.

Use this to support your reviewing process, not to replace it. Only relying on Claude as a reviewer does not count as a proper review. Also refrain from letting Claude publish reviews autonomously, in accordance with this post on showing AI output to people.

Build understanding

AI is effective at answering cross-cutting questions about the codebase that would otherwise require reading through multiple packages and services manually.

Example prompt
Which services use K8up for backups and which don't? Explain why.
Example output

Services using K8up: Redis, MariaDB, Forgejo, Nextcloud. All four follow the same pattern in pkg/comp-functions/functions/common/backup/backup.go: create an XObjectBucket for S3 storage, generate a K8up repository password, create a k8upv1.Schedule, and mount a custom backup script via ConfigMap. These services have no built-in backup operator, so K8up (wrapping restic) provides generic pod-level backup to S3.

Services not using K8up:

  • PostgreSQL (StackGres/CNPG), both operators have native continuous WAL archiving and PITR, which is superior to file-level backup.

  • Keycloak, stateless app that delegates backup to its embedded VSHNPostgreSQL dependency.

  • MinIO/Garage, object storage is the backup destination itself; backing it up would be circular.

Key insight: K8up is the generic fallback for Helm-deployed apps without native backup operators. When an operator provides its own backup (StackGres, CNPG), that is preferred because it enables WAL-level PITR.

Plan implementations

AI can help plan feature implementations before writing code. Plugins like Superpowers add a structured workflow: clarify intent, explore codebase, propose approach.

Example prompt
I want to add a parameter to VSHNPostgreSQL that orders a pizza once the instance
has successfully provisioned. Use this as a reference:
https://blog.crossplane.io/providers-101-ordering-pizza-with-kubernetes-and-crossplane/
Guide me through how I could implement such a feature.
Example output

Claude fetches the blog post, explores existing types and composition functions, then highlights a key difference:

  • Blog describes a Crossplane Provider, a standalone controller making direct external API calls.

  • AppCat uses Composition Functions, gRPC pipeline steps that emit Kubernetes resources, no direct API calls.

  • So "ordering pizza" means emitting a Job or CronJob that calls the API, not calling it from the function.

Claude then asks clarifying questions (PoC vs. production?), proposes approaches, presents a design, and writes a design document, all before any code.

Debugging

AI can act as a rubber duck for debugging. Explaining a problem to Claude often helps clarify your own thinking, and it can suggest angles you haven’t considered.

Example prompt
VSHNPostgreSQL instances intermittently fail to become ready after restore.
The backup exists and the restore job completes, but the cluster never reaches
a running state. What could cause this?