Slide 1

Slide 1 text

Spec-Driven Development with AI Haim Michael Lecturer | Holon Institute of Technology (HIT)

Slide 2

Slide 2 text

Haim Michael

Slide 3

Slide 3 text

My name is Haim Michael. I live in Tel Aviv, and I am the CEO & founder of Zindell Technologies, Ltd. My passion lies in teaching and lifelong learning. I have been into programming since my childhood. I have extensive experience with Java (30+ yrs), JavaScript (30+ yrs), C++ (30+ yrs), C# (20+ yrs), TypeScript (12+ yrs), Python (15+ yrs), PHP (20+ yrs), Scala (16+ yrs), and Kotlin (12+ yrs). Teaching advanced computer science topics at the university. blog.lifemichael.com Tel-Aviv, Israel

Slide 4

Slide 4 text

Economics Books in Hebrew Software Development Books in English blog.lifemichael.com My Books

Slide 5

Slide 5 text

Courses (in Hebrew) @ https://academy.lifemichael.com/he More Than 25 Online Courses on Udemy blog.lifemichael.com My Courses

Slide 6

Slide 6 text

XtremeJ XtremePython XtremeAI xtremej.dev xtremepython.dev meetup.com/lifemichael xtremeai.dev More Than 18000 Followers https://linkedin.com/in/lifemichael Kiro Ambassador XtremeJS xtremejs.dev More Than 2800 Video Clips youtube.com/lifemichael blog.lifemichael.com Conferences, Meetups & Community

Slide 7

Slide 7 text

Jacado - More Than 200 Games for Mobile Telephones – Years 2001- 2008 Zindell - Building AI Aligned Companies – Years 2026 - … nana.events prompo.ai zindrex.com ngager.pro jacado.games blog.lifemichael.com Passion to Build

Slide 8

Slide 8 text

The English Programming Language

Slide 9

Slide 9 text

The End of Code First Development The large-scale code generation shifts the bottleneck from code generation to understanding and control.

Slide 10

Slide 10 text

The Vibe-Coding Problem The prompts cannot be managed or traced. When the systems evolve, intent is lost, and the behavior becomes difficult to explain or reproduce.

Slide 11

Slide 11 text

Spec-Driven Development

Slide 12

Slide 12 text

Specs as the Missing Layer Specs capture what the system must do, what it must never do, and under which conditions behavior changes. They translate human intent into a form that humans and machines can review, version, validate, and share. Both by humans and by machines.

Slide 13

Slide 13 text

Spec-Driven Development AI Tools Created with ChatGPT. The prompt: Create a table that compares all the AI tools that support Spec-Driven Development.

Slide 14

Slide 14 text

The Kiro IDE

Slide 15

Slide 15 text

The Kiro IDE The Kiro IDE was developed by Amazon and is based on VS Code. Kiro is an Agentic IDE.

Slide 16

Slide 16 text

The Kiro Tools Family Kiro enables agentic development across multiple environments: Kiro IDE for visual development, Kiro CLI for terminal-based workflows, Kiro Mobile for managing agent sessions on the go, and Kiro Crew for persistent, long-running autonomous tasks.

Slide 17

Slide 17 text

The Development Process Steering Specs Documents Mermaid Diagrams Execution Requirements Design Phase List of Tasks EARS Easy Approach to Requirements Syntax The tasks are performed by the agents

Slide 18

Slide 18 text

AI Agents Hooks

Slide 19

Slide 19 text

What are Agent Hooks? Agent hooks are automated triggers that execute predefined agent prompts or shell commands when specific events occur in your IDE. Rather than manually asking for routine tasks to be performed, hooks set up automated responses to events such as: Linting Updating Other Files Create Tests Static Analysis Tools Code Formatting Updating Related Files Generating Documentation

Slide 20

Slide 20 text

Agent Hooks There are two ways to create new hooks. Either you create them manually or you ask Kiro chatbot to create them for you.

Slide 21

Slide 21 text

Creating Simple Hook

Slide 22

Slide 22 text

Creating Simple Hook

Slide 23

Slide 23 text

Creating Simple Hook

Slide 24

Slide 24 text

Subagents

Slide 25

Slide 25 text

What are Subagents? The subagents allow Kiro to run multiple tasks parallelly or to delegate specific tasks to subagents that specialize in those tasks.

Slide 26

Slide 26 text

Executing Subagents Manually We can launch subagents manually by instructing Kiro to do so via a prompt such as ”run subagents to do this and that...".

Slide 27

Slide 27 text

Executing Subagents Manually 1st Agent 2nd Agent

Slide 28

Slide 28 text

Orchestration Patterns

Slide 29

Slide 29 text

Common Orchestration Patterns There are common scenarios and common patterns when developing software applications with the help of AI agents: Reviewer Loop One agent produces work, and another verifies it. Event-Driven Agents External events trigger agents to perform specific tasks.

Slide 30

Slide 30 text

Common Orchestration Patterns Coordinator & Workers One agent divides the goal and combines the results. Sequential Pipeline Each agent (specialist) passes its output to the next agent. Parallel Specialists Several (specialist) agents that simultaneously do their work.

Slide 31

Slide 31 text

Tasks in Parallel

Slide 32

Slide 32 text

Running Tasks in Parallel When we click "Run all Tasks" on a spec, Kiro analyzes the task list and figures out which tasks can run at the same time, and builds a dependency graph from the task list.

Slide 33

Slide 33 text

Running Tasks in Parallel Taken from https://kiro.dev/blog/faster-smarter-specs/

Slide 34

Slide 34 text

Kiro Workflows

Slide 35

Slide 35 text

What is Kiro Workflow? Kiro Workflows automate complex, multi-step tasks by orchestrating multiple AI agents through sequential steps, parallel execution, and iterative loops. Each agent operates in an independent session, enabling reusable workflows that run with minimal human supervision.

Slide 36

Slide 36 text

What is Kiro Workflow? Was created using AI based on the diagram at https://kiro.dev/blog/introducing-workflows

Slide 37

Slide 37 text

What is Kiro Workflow? { "name": "deliver-change", "inputs": { "task": "prompt", "workdir": "string" }, "steps": [ { "type": "step", ”id": "plan", "agent": "wf-planner", … }, { ... } ] } Was created using AI based on the diagram at https://kiro.dev/blog/introducing-workflows

Slide 38

Slide 38 text

Cloud Based AI Agents

Slide 39

Slide 39 text

AI Agents in the Cloud Cloud-based AI agents can operate continuously without relying on a developer’s local machine. They can respond to events, execute longrunning tasks, collaborate with other agents, and scale dynamically while maintaining secure access to development tools and organizational resources.

Slide 40

Slide 40 text

Common Use Cases There are various cases in which having AI agents on the cloud can be highly useful: • Monitoring Systems • Responding to Incidents • Reviewing Code

Slide 41

Slide 41 text

Common Use Cases • Processing Data • Generating Recurring Reports • Handling Support Requests • Complex Development Tasks • Running Tests

Slide 42

Slide 42 text

Continuous Orchestration

Slide 43

Slide 43 text

Persistent Orchestration Kiro Crew supports long-running work across multiple sessions and allows us to schedule specific tasks.

Slide 44

Slide 44 text

Agents Observability Agent orchestration must remain observable. Kiro Crew exposes agent activity, tool calls, approvals, and results as the work progresses.

Slide 45

Slide 45 text

Best Practices

Slide 46

Slide 46 text

Update Spec First Whenever you add a new feature or fix a bug, start by updating the spec. The code, including the tests, should be aligned with the spec.

Slide 47

Slide 47 text

Detailed Steering Documents When having detailed steering documents, the steps that follow will produce better results. By doing so, you reduce re-work, backtracking, and misalignment across humans and the AI.

Slide 48

Slide 48 text

Technical Design When we finish with the requirements, Kiro produces the design artifacts: diagrams, interfaces, schemas, and API endpoints. Make sure you carefully review these documents. Introducing AI-driven changes might pose technical challenges you will need to overcome.

Slide 49

Slide 49 text

AI Optimization Make sure to configure the relevant MCP servers and to install the skills and powers that will assist AI. Powers bundle MCP tools with knowledge and workflows. They activate dynamically based on context. Skills are portable instruction packages that bundle instructions, scripts, and templates, and Kiro can activate them when relevant to your task.

Slide 50

Slide 50 text

The Right Level of Orchestration Effective orchestration is becoming more complex. Parallelism does not guarantee better results.

Slide 51

Slide 51 text

The Takeaways

Slide 52

Slide 52 text

Coding is Cheap AI dramatically reduces the cost of generating code. The real challenges now lie in defining intent, controlling execution, verifying results, and preserving knowledge over time.

Slide 53

Slide 53 text

Specs Coordinate The Development Specs provide agents with shared requirements, constraints, design decisions, and acceptance criteria. They allow humans and multiple agents to work toward the same outcome.

Slide 54

Slide 54 text

Higher Level of Abstraction Developers increasingly define goals, architecture, constraints, and quality standards while AI agents perform most of the coding. The developer becomes the orchestrator of software creation.

Slide 55

Slide 55 text

Q&A nana.events Thanks for attending my talk :) nana.events WhatsApp +972.54.6655837 prompo.ai [email protected] https://blog.lifemichael.com prompo.ai ngager.pro life michael zindrex.com XtremeJ xtremej.dev life michael XtremeJS xtremejs.dev XtremePython xtremepython.dev XtremeAI xtremeai.dev lifemichael.com meetup.com/lifemichael Monolith vs Microservices