Vibe Coding hat sich verändert

Summary

The speaker argues that AI-assisted software development has changed rapidly, making some once-common practices less necessary. As newer models become more autonomous, developers may get better results by giving them a clear goal instead of detailed, step-by-step instructions.

Key themes

• Less prompting, more autonomy: The speaker demonstrates “Screen Forge,” a Screen Studio-style clone that Claude reportedly built after receiving a broad, one-line instruction and working for about 24 hours. The model chose the architecture and produced editable timelines and working zoom effects without detailed guidance.
• Longer, more independent agent sessions: Improved context handling and automatic summarization let agents work for extended periods without developers manually dividing tasks into loops or frequently restarting sessions. The speaker still recommends separate sessions for unrelated tasks or discussions with conflicting instructions.
• Fewer, more targeted repository instructions: The speaker discusses an ETH Zurich paper that reportedly found AGENTS.md-style context files did not improve agent success rates overall and increased inference costs by 20%. He recommends keeping only instructions for genuinely unusual project requirements. For routine requirements, he favors deterministic scripts, hooks, lint rules, and CI checks over prose instructions.
• Planning and architecture: Human planning remains important for deciding what a product should—and should not—do. However, the speaker says AI systems are increasingly capable of making architectural decisions, and developers should avoid assuming that advice from only a few months ago still applies.
• Parallel agents: Agents can increasingly launch sub-agents themselves, reducing the need for people to assign separate “frontend” and “backend” agents manually. Parallel work can still help when tasks are clearly separated, but elaborate agent setups may be more showy than efficient.
• Code review, security, and responsibility: The speaker expects human review to become more selective rather than requiring someone to inspect every line. He says leading models can write strong code and may catch security issues, but recommends explicitly requesting security reviews for important features. Humans still set priorities, decide how much effort to spend, and remain accountable for outcomes; the appropriate level of caution depends on the software’s risk.
• Changing roles for IT professionals: Developers may increasingly act as product owners, project leads, or decision-makers rather than writing every part of the code. The speaker also highlights advisory and consulting work, particularly helping organizations modernize legacy systems and integrate software with existing hardware, processes, and staff. He argues that AI does not eliminate the need for computer science expertise, but makes adaptability more important.
• Broader concerns: The speaker acknowledges open questions about how delegating more thinking to AI may affect human skills and curiosity, while maintaining that AI is now substantially better at many concrete programming tasks.

Guides, Demonstrations, and Sources Mentioned

• Demonstration: Claude’s recreation of Screen Studio as “Screen Forge.”
• Tutorials: Matt Pocock’s tutorials on context management and agent workflows; the speaker notes that some advice may now be outdated.
• Example: Matthew Berman’s video showing Codex working on a task for several days.
• Research: An ETH Zurich paper evaluating AGENTS.md files. The speaker summarizes its findings as no general improvement in success rates and 20% higher inference costs.
• Industry example: MSG’s modernization of a decades-old software system used in the long-distance transport of seamless pipes, illustrating the complexity of real-world digital transformation.

Main Speaker and Sources

• Niklas Steenfatt: The presenter and main speaker, drawing on his experience as a computer scientist and former professional developer at Meta.
• Sources and examples cited: ChatGPT’s comparison of past and current advice, Anthropic and Claude, the ETH Zurich paper, Matt Pocock, Matthew Berman, and MSG’s industrial modernization example.