Vibe Coding for Production Porto Tech Hub 2025.pptx

The presentation explores usage of Agentic AI platforms (often "vibe coding") as an abstraction layer and new approach for software development: highlighting agentic coding platforms, prompt-to-app workflows, and the shift toward specification-driven, AI-first methodologies. Here we are exploring where such approaches can work for for enterprise workflows and enterprise applications.
It emphasizes practical lessons from real-world projects, advocating for smarter tooling, robust documentation, and a developer experience where humans focus on architecture and critical review while AI handles much of the coding if its specified enough.

1. Vibe Coding for Production AI agent platforms for enterprise development Porto Tech Hub 2025

2. Speaker 2 Bio Emergn AI Lab Technical architect for AI+ML, data enabled solutions Experimentation, rapid prototyping Innovation & feasibility reality checks Python, Azure, Power BI, JS/TS etc. Muntis Rudzitis LEAD DATA SCIENTIST, EMERGN

3. Vibe coding Agentic coding platforms Practices & Methodologies AI-First approach Our experience Closing thoughts

4. What is vibe coding? Using AI as an abstraction layer between developer and the code Giving AI larger tasks Little or no reviewing Disposable outcome PoC 4

5. Agentic coding 5 AI platforms (low/no code) Coding tools (full code)

6. The dream of prompt to app 6 The Dream The Reality https://adamtheautomator.com/vibe-coding-with-ai/

7. The dream of prompt to app 7 Launched Beautiful portfolio page with nice UI Tool / template pages ToDo / Focus apps Small helper applications Niche specialized applications variants More complex apps migrate to full code eventually

8. What works out of the box Exploration Test an idea Great to oneshot multiple variants Great for MVP / 70% Full package for solo development Some workflow and UX changes Modular applications ( flat and wide ) Limited context within module Visual and typical CRUD operations 8 Hard to edit Large edits with huge minimize code impact Guesses when to do bigger refactor Non-existent / weak / complex collaboration Surgical edits with minimal impact Complex, partially reusable modules AI mises the context -> reimplements “Invisible” or complex logic What works What doesn’t

9. Practices & Methodologies 4 methods to adapt for enterprise application development

10. 1. Prompt quality still matters General prompting guidelines Details Specification Plan Prompt “frameworks” (C.L.E.A.R etc.) Platform or tool specific tips and features Platforms want you to succeed Meta-prompting Zero-shot and few-shot approaches 10

11. 2. Vision Documentation PRD (product requirements document) Vision & Goals High level user stories / functionality Design / UI User description (roles, problems) Helps AI to hallucinate the correct way Tools: Any AI Chat ( openai , gemini , anthropic, grok) ChatPRD.ai Agents (copilot/cursor) on existing code bases Works as a persistent memory for AI 11

12. 3. README for AI AGENTS . md README for agents Most tools check for these automatically Extra instructions about: Environment setup How how to.. create / run / test Nested files per module, project, etc. Keep these updated! Similar trends Agent-friendly documentation ReadMe.llm Works as a persistent instruction for the AI 12

13. 4. Methodologies Specification driven development (SDD) Test driven development (TDD) Defined workflow Specification Architecture & Plan Tasks & Implementation Tools wins over models Spec-Kit by GitHub (OSS tool) Kiro by AWS (IDE) OpenSpec (OSS CLI) BMAD-method Works as a persistent specification to continuously cross reference for impact and changes 13

14. Our experience By building products and prototypes 14

15. DECISION CHECKPOINT AI-First Approach 15 Establish business
expectations AI ADVISORY Co-create functional
prototype FUCNTIONAL PROTOTYPING Evaluate prototype
and decide next steps Implement AI-first framework DEVELOPMENT PROCESS SETUP Develop production- ready product PRODUCT DEVELOPMENT Enhance and develop features PRODUCT IMPROVEMENTS PRD AGENTS.md and Spec PRD → Spec Tech stack Automation and guardrails BA + SME UX + SME Follows SDLC

16. Our experience Hybrid - AI development platforms + full code Feature Complexity Estimation Low complexity (UI, CRUD, reuse) → AI platform High complexity (invisible logic, security) → Full-code Process Adaptations Low complexity → just implement Higher complexity → discussing in-depth technical plans Experimenting with starting with DB / backend first, UI with AI platform Use AI for UI elements, intentionally mock backend Code review are more “per module” than “per PR” Document architecture for AI reuse (!!!) 16 Collaboration Shift Sync for humans, trace decisions = knowledge for AI Meeting transcripts, code summaries, and architectural notes feed and help the AI Experimenting with snapshot branches to help to scope and review AI-generated code (code reviews are still difficult) Pair prompting Treat AI like a junior dev (or 10) - needs clear, detailed instructions with limited scope and reference to good examples Learn git

17. Closing thoughts Private and confidential 17

18. Where is this going? Smarter Tooling → Smarter Instructions → Better Code Platforms guide developers toward better prompting and planning Architect mode, implementation plans, structured ToDos/Specs Vibe coding as a “glue” Learn what's good – architecture and code patterns Specification-first development Developers are shifting from "just build" to "build with guardrails" Risk-aware planning and architectural notes are essential for reliable outcomes Frameworks having specialized AI variants Opinionated platforms will win Framework documentation for AI Agentic tooling >>> LLM provider 18 Developer Experience (DX) Shift Experienced developers benefit most: Know what “good” looks like Has experience delegating Quality linked to good local patterns Architecture flows: old: v1 → peer review → feedback loop → implementation new: AI v1 → review v2 → peer review → implementation Ownership Shift Less attachment to code → more critical reviews Initial versions are bigger, often include “bells and whistles” Redoing and experimentation are cheap

19. Muntis Rudzitis [email protected] www.linkedin.com /in/ muntis-rudzitis