Prompt. Build. Repeat. – The Vibe Coding Mindset | NDC Copenhagen 2025

This talk explores how Vibe Coding works in practice and how it’s transforming the everyday workflows of people in software development. Tools like GitHub Copilot, Lovable, and Cline enable a new kind of collaboration between humans and AI, where prompting becomes the core skill. Instead of focusing on syntax and boilerplate, developers can concentrate on creativity, goals, and iteration.

But Vibe Coding is about more than just speed. It lowers the barrier to entry, democratizes technical capabilities, and unlocks entirely new creative possibilities. At the same time, it raises important questions around data privacy, quality assurance, and the evolving role of tech professionals in a world where anyone can generate working code.

Whether you’re already experimenting with AI or just curious about what’s next – this talk will provide inspiration, strategies, and a fresh perspective on what software development is becoming: Prompt. Build. Repeat.

Transcript

1. Prompt. Build. Repeat. The Vibe Coding Mindset Daniel Sogl daniel.sogl@thinktecture.com Software architect
2. Don’t do this in production I want to have a web app that helps me generate social media posts. I should be able to authenticate myself to securely access my account. Once I’m logged in, I want to enter a topic into a free-text field. The app should then fetch current sources related to that topic using the Perplexity API. Based on these sources, it should use the ChatGPT API to generate relevant, engaging, and reach-optimized social media posts. All generated posts should be persistently saved in a database and linked to my authenticated user account. Let's vibe code How to AI Coding tools boost your productivity Code smarter, not harder
3. Daniel Sogl • Software architect @ Thinktecture AG • MVP – Developer & Web Technologies • Focus: Angular and Generative AI • Socials: https://linktr.ee/daniel_sogl About me The Vibe Coding Mindset Prompt. Build. Repeat.
4. A strategic approach to AI-assisted development Our Journey: • Discover why experienced developers struggle with AI tools • Build sustainable AI coding practices that work • Master the strategic setup that 95% of teams skip What We Won't Do: • Fall for vendor productivity myths without evidence • Ignore the growing frustration with current AI approaches • Pretend vibe coding works for production systems Prompt. Build. Repeat. The Vibe Coding Mindset What We'll Explore Together
5. We're Living Through a Historic Moment • Cursor: $2.5B valuation, exponential growth • GitHub Copilot: 50,000+ organizations • Up to 40% of code is now AI-generated • Path to "a billion programmers” - GitHub CEO Prompt. Build. Repeat. The Vibe Coding Mindset The AI Coding Revolution https://www.elitebrains.com/blog/aI-generated-code-statistics-2025 https://www.netcorpsoftwaredevelopment.com/blog/ai-generated-code-statistics https://blog.github.com
6. From Andrej Karpathy Tweet to Cultural Phenomenon • February 2nd, 2025: Karpathy coined the term • Now in Webster's Dictionary • Meme status: "Vibe coding is easy. Vibe debugging is the hard part" Prompt. Build. Repeat. The Vibe Coding Mindset Enter "Vibe Coding"
7. Not All AI Coding Is Vibe Coding • AI Typing Assistant: "Complete this for loop” • AI Pair Programming: "Help me debug this authentication issue” • True Vibe Coding: "Build me an entire user management system" Prompt. Build. Repeat. The Vibe Coding Mindset The Vibe Spectrum
8. Vendor Claims vs Reality The Marketing: • GitHub Copilot: 55% faster task completion • Amazon CodeWhisperer: 57% speed improvements • Cursor users report: 5-10x productivity multipliers • 96% of developers use Copilot the same day they install it Prompt. Build. Repeat. The Vibe Coding Mindset The Promise Machine https://github.blog https://aws.amazon.com/q/developer/ https://cursor.com/students
9. • Lovable: $0 → $100M ARR in 8 months • Cursor: $500M ARR, $2.5B valuation • GitHub Copilot: $400M ARR, 281% growth BUT... • 95% of AI pilots fail to achieve revenue acceleration • 42% of companies abandoned most of their AI initiatives in 2025 • 19% slower coding despite feeling 20% faster Prompt. Build. Repeat. The Vibe Coding Mindset $100M Paradox https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf https://beam.ai/agentic-insights/why-42-of-ai-projects-show-zero-roi-(and-how-to-be-in-the-58-) https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
10. Controlled Study: 16 Experienced Developers, 246 Real Tasks • Actual performance: 19% SLOWER with AI tools • Developer perception: Believed they were 20% faster • Perception gap: 40 percentage points difference • Tool used: Cursor with Claude Sonnet 3.5 Prompt. Build. Repeat. The Vibe Coding Mindset Reality Check - The METR Study https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
11. • METR study: 16 experienced developers, controlled environment = reliable data • Vendor claims: Mixed skill levels, cherry-picked scenarios = marketing • The productivity gap exists even among experts when using current methods • Implication: The problem isn't user skill - it's methodology Prompt. Build. Repeat. The Vibe Coding Mindset Why The Studies Matter - Controlled vs. Anecdotal
12. GitClear Analysis: 153-211 Million Lines of Code • Code churn: Projected to double in 2025 • Copy-pasted code: 8.3% → 12.3% (+48%) • Code refactoring: 25% → <10% (-60%) • Delivery stability: 7.2% decrease (Google DORA Report) Prompt. Build. Repeat. The Vibe Coding Mindset Code Quality Catastrophe https://www.gitclear.com/ai_assistant_code_quality_2025_research
13. Veracode 2025 Study: AI Code Security Failures • 45% of AI-generated code fails security tests • Java security failure rate: 72% • Secret leakage: 40% higher in Copilot repos Prompt. Build. Repeat. The Vibe Coding Mindset Security Nightmare https://www.veracode.com/blog/genai-code-security-report/ https://blog.gitguardian.com/yes-github-copilot-can-leak-secrets/ https://snyk.io/reports/ai-code-security/
14. The methodology gap that's killing productivity • The tools work for simple, isolated tasks • They fail in complex, real-world codebases because they lack context • Solution: Stop "vibe coding" and start strategic AI integration • Next: How to give AI the context it needs to succeed Prompt. Build. Repeat. The Vibe Coding Mindset The Real Problem Isn't AI - It's How We Use It
15. Why coding tools fail to support you • LLMs were trained on older data • LLMs don’t understand your architecture without help • LLMs don’t know your company domains • LLMs don’t know your coding standards • Developers don’t write prompts describing every edge case or needed context to solve tasks • Different tasks require different context Prompt. Build. Repeat. The Vibe Coding Mindset Limitations of LLMs
16. Simplify your prompts • Project-specific coding standards • Architectural patterns and constraints • Domain knowledge embedding • Step-by-step reasoning requirements How to generate?: • Ask the AI to analyze your codebase • Use predefined instructions • Itterateive update your instructions Prompt. Build. Repeat. The Vibe Coding Mindset Fighting Back - Custom Instructions
17. AGENTS.md • The problem: each tool uses different file names for custom instructions • AGENTS.md aims for becoming supported by all AI tools • Supported by: Cursor, GitHub Copilot, Aider and more Prompt. Build. Repeat. The Vibe Coding Mindset Standardization of Custom Instructions
18. Prompt. Build. Repeat. The Vibe Coding Mindset
19. The key for ”intelligent” workflows • Open-source protocol developed by Anthropic • Provides a consistent way for LLMs to interact with external resources • Client-Server architecture: AI applications (clients) request context from external services (servers) • Official servers are available for GitHub, Atlassian, Playwright, Stripe, Databases and more • It’s the key to useful AI-coding setups in complex environments Prompt. Build. Repeat. The Vibe Coding Mindset Model Context Protocol (MCP)
20. Prompt. Build. Repeat. The Vibe Coding Mindset MCP workflows Developer asks: “Explain failing tests in PR #42.” Copilot calls GitHub MCP → fetches PR diff + CI log LLM reviews context → returns root-cause & fix steps Dev clicks “Apply fix” → Copilot edits code & opens new PR
21. Prompt. Build. Repeat. The Vibe Coding Mindset
22. • Not all LLMs are equal: Performance varies widely • Understand reasoning vs. non-reasoning models • Context size matters greatly in coding scenarios • Benchmarks help guide practical model choice • Cost per 1 million input/output tokens Prompt. Build. Repeat. The Vibe Coding Mindset Choosing the right LLM for your tasks
23. Choose wisely Prompt. Build. Repeat. The Vibe Coding Mindset Popular LLMs for software development Model Provider Reasoning Context Window SWE-bench $ / 1M tokens Claude 4 Opus Anthropic 200k tokens 72.5 % $15 / $75 Claude 4 Sonnet Anthropic 200k tokens 72.7 % $3 / $15 GPT-5 (high) OpenAI 128k tokens 74.9 % $1.25 / $10 GPT-o3-mini (high) OpenAI 200k tokens 49.3 % $1.10 / $4.40 Gemini 2.5 Pro Google 1M tokens 63 % $1.25 / $10 DeepSeek V3 DeepSeek 163k tokens 38.3 % $0.25 / $0.85 Grok 4 xAI 256k tokens 72 % $3 / $15 Kimi K2 Moonshot 131k tokens 65.8 % $0.13 / $0.13
24. CO-STAR Framework for Prompting • Context: Project background, constraints • Objective: Specific goals, success criteria • Style: Coding patterns, conventions • Tone: Collaborative, step-by-step • Audience: Experience level, domain • Response: Format, structure requirements Prompt. Build. Repeat. The Vibe Coding Mindset Prompt Engineering That Works
25. Proven Tools Stay Essential • Static Analysis Foundation: ESLint, SonarQube, SAST tools remain critical • AI-Enhanced Review: CodeRabbit, Copilot suggestions as additional layer • Automated Security Scanning: Snyk, Veracode for vulnerability detection • Comprehensive Testing: Unit, integration, security tests (AI can generate, not replace) • Human Oversight: Architectural decisions, business logic validation Prompt. Build. Repeat. The Vibe Coding Mindset Defense in Depth: Quality Gates for AI-Enhanced Development
26. How to Do AI Coding Right Setup: • Define custom instructions/rules files • Configure MCP servers for your stack • Choose task-appropriate LLMs Process: • Use CO-STAR prompt framework • Require AI to explain before coding • Implement automated testing generation • Maintain human review of critical paths Quality: • Run security scans on all AI code • Code quality: Cyclomatic complexity, test coverage, refactoring frequency • Track technical debt metrics Prompt. Build. Repeat. The Vibe Coding Mindset Best Practices Checklist
27. Prediction vs Preparation Inevitable Trends: • AI handles increasingly complex tasks • Autonomous agents are already here: Claude Code, Cursor, v0.dev, Replit Agent • Context integration is accelerating: MCP adoption, RAG improvements • Quality tooling is emerging: AI code review, automated security scanning Critical Decisions Today: • How do we maintain code quality? • What skills should humans develop? • How do we ensure security and privacy? Prompt. Build. Repeat. The Vibe Coding Mindset The Future We're Building
28. Your Next Steps 1. Experiment Responsibly • Start with low-risk projects • Implement quality safeguards first • Measure Everything Track actual productivity (not perception) • Monitor code quality metrics • Assess security vulnerabilities 1. Invest in Skills • Prompt engineering techniques • AI tool orchestration • Quality assurance processes Prompt. Build. Repeat. The Vibe Coding Mindset Call to Action
29. Prompt. Build. Repeat. The Vibe Coding Mindset The best developers won’t be replaced by AI, they’ll be the ones who learn to wield it
30. Thanks for your attention Daniel Sogl @sogldaniel daniel.sogl@thinktecture.com