Status Quo of AI-based Software Development

Softwareallianz Chapter Meeting Stuttgart

Frederik Wystup

MeiLuft GmbH & Co. KG

What's This About?

  • AI coding tools evolve monthly
  • Central question: How do we deploy these strategically?
  • From idea to app in minutes - but professionally
  • Context engineering is more critical than code itself
  • AI as "junior developer" in the team

About Me

Frederik Wystup

  • Engineer & Entrepreneur
  • Co-founder MeiLuft GmbH & Co. KG
  • Academic Supervisor DHBW Mosbach
  • 25+ years IT experience

Focus Areas

  • Electron-optical sensors
  • AI for control & image analysis
  • AI-assisted development
  • Open-source AI coding tools

Typical Challenges

You might be familiar with...

  • Expensive rework in software projects
  • Team scaling - The right people at the right time
  • Integrating new technologies - AI, Cloud, DevOps
  • Cost control with increasing complexity
  • Speed - Time-to-market vs. quality
  • Expectations - "Can't we just build this with ChatGPT?"

→ AI coding can help here - if used correctly

The Answer to Everything...

42

Without context you only get "42"

Why Context is More Important Than Code

Context Engineering = The art of giving AI the right context

  • Project structure & architecture decisions
  • Coding standards & best practices
  • Testing requirements (TDD, coverage, E2E)
  • Security & quality gates
  • Business domain & domain logic

→ Context documentation is more valuable than generated code!

The Evolution of AI Coding

Stage 1: Vibe Coding

ChatGPT
Gemini
Grok (X.AI)
MS Copilot
  • Pure browser-based, no setup needed
  • Fast & easy - code via copy & paste
  • Good for prototypes & experiments
  • But: "App in 5 minutes" ≠ production ready

"Why is software so expensive? You can build everything in 5 minutes now..."

→ The question your customers ask

The Evolution of AI Coding

Stage 2: Integrated Platforms

bolt.new
Google AI Studio
Replit
  • Integrated development environment in browser
  • Direct preview & deployment
  • Still primarily "vibe coding"
  • Limitation: Vendor lock-in, little control

The Evolution of AI Coding

Stage 3: Agentic Coding 🎯

Claude Code
Gemini CLI
GitHub Copilot Workspace
  • Tool access: AI can read, write, test, commit files
  • Context-aware: Understands project structure & dependencies
  • Iterative: Plan → Implement → Test → Fix cycles
  • Governance: Hooks, guards, human-in-the-loop

→ This is professional AI-assisted development

Avoiding Vendor Lock-in

Multi-tool strategy for professional development

⚠️ The Problem

  • Claude Code is currently very good - but everything depends on Anthropic
  • Risks: Server outages, API changes, pricing changes, availability
  • Dependency on a single provider = business risk

✅ Multi-Tool Approach

  • Claude Code, Gemini CLI
  • Continue, Cline, Aider
  • OpenAI Codex
  • Always have alternatives ready

🖥️ Local LLMs

  • Qwen, DeepSeek, Llama
  • Almost as good as cloud LLMs
  • Full control & privacy
  • GPU hardware needed (on-prem or rented)

DEMOXaresAICoder

Open-source tool for professional AI coding sandboxes

At the push of a button:

  • Isolated Docker development environment
  • Pre-installed tools (Git, Node, Python, etc.)
  • Claude Code / Gemini CLI integration
  • Secure sandbox for experiments

github.com/dg1001/xaresaicoder

DEMOhn-gems Project

Evolving existing projects with AI

Task: Add German audio version to existing podcast

  • Load project in XaresAICoder
  • Provide Claude Code with context
  • Implementation: Text-to-speech for German
  • Testing & iteration
  • Git commit with AI-generated message

github.com/DG1001/hn-gems

BMAD-METHOD

"Breakthrough Method for Agile AI Driven Development"

19+ Agents

Specialized AI personas

50+ Workflows

Guided processes

20.8k ⭐

GitHub Stars

Scale-Adaptive Intelligence

  • Quick Flow: Bug fixes, small features
  • BMad Method: Products, platforms (PRD + Architecture + UX)
  • Enterprise: Compliance, security, DevOps

github.com/bmad-code-org/BMAD-METHOD

Ralph - The While-Loop Technique

Iterative AI coding automation

The Concept

while :; do cat PROMPT.md | npx --yes @sourcegraph/amp ; done

Continuous iteration: Prompt → Execute → Learn → Improve

✅ Use Cases

  • Develop prototypes
  • Start greenfield projects
  • Learn new programming languages
  • Experiment with new technologies

💡 Principle

  • Deterministic errors
  • Iterative prompt refinement
  • Like tuning a guitar: Step by step better
  • Automation of repetitive tasks

ghuntley.com/ralph/

What Do Studies Say About AI Coding?

Important: Most studies examine "vibe coding" (autocomplete), not "agentic coding"

✅ Positive Effects

  • GitHub/Accenture 2024: 73% stay in flow better
  • 87% save mental energy on repetitive tasks
  • Harness: 10.6% more PRs, 2.4% faster cycles
  • 60-75% feel more fulfilled at work

⚠️ Critical Findings

  • GitClear 2025: 41% higher churn rate (more rework needed)
  • Medium 2024: 19% slower, but believe 20% faster!
  • NYU: 40% of AI code with security issues
  • Perception gap: Felt ≠ actual productivity

Conclusion: AI coding helps with flow & mental relief, but code quality and actual speed must be critically measured. Therefore, using agentic coding with governance is important, which can deliver significantly better results - but long-term studies are still missing.

Agentic Governance

Define context before AI starts

  • Project context: Architecture, tech stack, patterns
  • Company context: Standards, security requirements, compliance
  • Development context:
    • Testing: TDD, E2E, coverage goals
    • Quality gates: Linting, type checking, reviews
    • Security: OWASP Top 10, dependency scanning
  • Review process: AI-assisted, but human approved

→ Don't rely 100% on AI. Human remains architect & quality gate.

The Right Expectations

AI = Junior team member (without domain knowledge)

What AI can do ✓

  • Generate boilerplate code
  • Perform refactoring
  • Write tests
  • Create documentation
  • Fix bugs (with context)
  • Iterate quickly

What AI cannot do ✗

  • Architecture decisions
  • Understand business domain
  • Debug complex bugs
  • Evaluate trade-offs
  • Recognize security risks
  • Think long-term

Important: AI doesn't replace the junior, but complements the team!
Senior developers can focus on high-value tasks.

Best Practices for AI Development

✅ Do's

  • Context first: Detailed project description
  • Develop iteratively: Small steps, frequent feedback
  • Automate testing: TDD with AI support
  • Code reviews: Human + AI
  • Documentation: Document prompts and decisions
  • Experimentation: Create prototypes quickly

❌ Don'ts

  • Blind trust: Always check AI results
  • Without context: No detailed requirements
  • Direct to production: Without testing/QA
  • Ignore security: AI doesn't know all risks
  • Domain knowledge: AI has no domain expertise
  • Underestimate complexity: Real projects need structure

Key Takeaways

  1. Context > Code - Documentation of prompts & context is more valuable than generated code
  2. Understand evolution - From vibe coding to agentic coding for production quality
  3. Use tools - XaresAICoder, Claude Code & Co for professional workflows
  4. Establish governance - Define context, tests, quality gates BEFORE coding
  5. Right expectation - AI as junior team member, not as developer replacement
  6. Experiment - Use sandbox environments to safely experiment

→ Prepare teams for the "Agentic Age"

Questions & Discussion

Let's discuss!

Frederik Wystup

MeiLuft GmbH & Co. KG

GitHub: github.com/dg1001
XaresAICoder: github.com/dg1001/xaresaicoder
hn-gems: github.com/DG1001/hn-gems

Thank You!

Have fun experimenting with AI coding!

Softwareallianz Stuttgart Chapter Meeting
November 2025