---
title: "AI Agents in Practice"
description: "AI Agents in Practice — a 5-part article series on the SaaSForge blog."
url: https://saasforge.cz/series/ai-agents-in-practice/
language: en
translation_cs: https://saasforge.cz/cs/series/ai-agents-in-practice/
---

# AI Agents in Practice

AI Agents in Practice — a 5-part article series on the SaaSForge blog.

1. [How We Use Claude Code Skills to Codify Our Development Patterns](https://saasforge.cz/blog/claude-code-skills-guide/) (2026-04-16): A practical guide to teaching your AI coding assistant your codebase conventions using Claude Code Skills — reusable Markdown templates that prevent repeated code review corrections.
2. [140 Pull Requests in a Week: Orchestrating Claude Code Agents at a Hackathon](https://saasforge.cz/blog/claude-code-agents-hackathon/) (2026-06-12): Lessons from a hackathon that capped off a months-long service renovation with massive AI assistance — how to run Claude Code agents in parallel, the keyboard shortcuts that keep you sane, what survives a reload, and a status line that keeps you oriented.
3. [When Your AI Agent Stops Trusting the Language Server: Three Bugs Behind the Noise](https://saasforge.cz/blog/jdtls-gradle9-lsp-noise-ai-agents/) (2026-06-19): How a Gradle 9 incompatibility in jdt.ls, a missing Lombok javaagent, and a 120-second timeout taught our coding agent to ignore diagnostics — and how we fixed all three.
4. [Architecture Tests as a Seatbelt Against Agentic Drift](https://saasforge.cz/blog/archunit-agentic-drift/) (2026-06-26): When a tireless AI agent makes a hundred plausible-looking changes, what stops the architecture from quietly eroding? We wired ArchUnit into the CrushLog backend as an executable seatbelt — and the rules found bugs a human audit missed.
5. [Linting the Prompt Layer](https://saasforge.cz/blog/linting-the-prompt-layer/) (2026-09-11): We wrote our conventions in CLAUDE.md, then said prose doesn't fail a build. So we turned the prose into a skill with executable gates, pointed a linter at the skill, and open-sourced the result. Here's what the linter caught — including three broken images on this very site.
