Spec Driven Development | AI Summit Vilnius

Transcript

1. AI SUMMIT VILNIUS · 2026 Spec Driven Development The end of vibe coding. SPEAKER Daniel Sogl daniel.sogl@thinktecture.com Principal Consultant · Thinktecture AG
2. SPEAKER Daniel Sogl Principal Consultant at Thinktecture AG, focused on helping teams ship software faster and more reliably with AI. linktr.ee/daniel_sogl 01 · ROLE Principal Consultant Thinktecture AG — modernizing enterprise software across web, cloud, and AI. 02 · FOCUS Developer Productivity & Gen AI Spec-driven workflows, agentic coding, and everything that stops devs guessing what the machine just did. 03 · COMMUNITY MVP · Web Technologies Speaker, author, and contributor across conferences, open-source projects, and the modern-web ecosystem. Spec Driven Development The End Of Vibe Coding 2 / 27
3. Realistic workloads expose the gap. The "speed win" is often an illusion. Defects skew toward security. Sources: SWE-Bench Pro (Sep 2025) · Stack Overflow Developer Survey 2025 · CodeRabbit Report (Dec 2025) THE REALITY CHECK AI agents fail more than you think. 01 / PROBLEM Pass-rates collapse as soon as tasks leave toy-repo territory. Debugging agent output eats the time it claims to save. Agents confidently ship code with exploitable flaws. <50% SWE-bench Pro score 66% Devs spent more time fixing AI code 2.74× Security vulnerabilities vs. humans Spec Driven Development The End Of Vibe Coding 3 / 27
4. Give the same agent a complete, structured spec and all five of these weaknesses flip into strengths. That is the premise of the rest of this talk. WHY AGENTS STUMBLE Five core limitations of AI coding agents. 02 / LIMITS 01 Short memory Agents forget what you told them two turns ago — the more context, the more they lose. 02 No trade-off sense They copy patterns without understanding why. A choice that's right for one project may be wrong for yours. 03 Ripple blindness A change in one place breaks another. The agent can't see the domino effect across the project. 04 Missing company rules Doesn't know your company's unwritten rules — the tribal knowledge built over years. 05 Silent failures Disables safety checks, fakes data, and happily reports "all tests green". Spec Driven Development The End Of Vibe Coding 4 / 27
5. DEFINITION What is Spec‑Driven Development? A development methodology where a detailed, living specification drives the entire process — especially the collaboration with AI coding agents. INTENT ARTIFACTS CODE 03 / DEFINITION 1 Spec first Write the specification before any code. The spec is the source of truth — not the diff. 2 AI-optimized Specs hand AI agents the full context upfront: goal, constraints, acceptance. 3 Iterative Specs evolve with the project. They are not throwaway requirements docs. 4 Consistency Uniform architecture across features — the constitution keeps agents in lane. Spec Driven Development The End Of Vibe Coding 5 / 27
6. SDD isn't just a developer tool — it's a team communication protocol with AI as the executor. WHY THIS MATTERS TO YOU What SDD changes — for every role. 04 / ROLES Developer Specs replace endless prompts. You describe intent once — the agent works from a shared source of truth instead of guessing from fragments. Product Manager Review intent before code exists. Acceptance criteria become the first artifact — not an afterthought bolted on after the PR. QA & Testing Acceptance criteria are written on day one — in plain language. Test plans emerge from the spec, not from reverse-engineering the code. Manager & Leadership Traceable audit trail from intent to commit. Faster delivery, fewer rework cycles, and "show me the spec" replaces "why did we build this?" Spec Driven Development The End Of Vibe Coding 6 / 27
7. Treat AI agents like literal-minded pair programmers, not search engines. HOW SPECS CHANGE THE GAME Four mechanisms — how specs improve AI. 05 / MECHANISMS MECHANISM 01 Complete context The agent walks in with the full picture — goal, constraints, acceptance criteria — instead of assembling it from scratch on every turn. MECHANISM 02 Structured reasoning A spec forces explicit thinking: invariants, edge cases, error paths. Hallucinations drop because there's nothing left to invent. MECHANISM 03 Validation checkpoints Gates between phases — spec, plan, tasks, code — let humans review intent, not just the diff. The agent never sprints past the red flag. MECHANISM 04 Autonomous implementation Specs plus tests become guardrails. The agent can run for hours, self- validate, and still produce a reviewable trail. Spec Driven Development The End Of Vibe Coding 7 / 27
8. SDD SCOPE What & Why — intent level ARTIFACT Markdown spec, plan, tasks AUTHORED BY Human writes intent · AI drafts BDD SCOPE System behavior — feature level ARTIFACT Gherkin scenarios, step defs PATTERN Given / When / Then TDD SCOPE Code correctness — unit level ARTIFACT Unit and contract tests PATTERN Red / Green / Refactor SDD wraps BDD and TDD — it tells the agent what to build and why, so BDD scenarios and TDD tests drop out almost for free. POSITIONING How SDD relates to TDD and BDD. 06 / COMPARISON Spec-Driven Development Behavior-Driven Development Test-Driven Development Spec Driven Development The End Of Vibe Coding 8 / 27
9. → → → Each arrow is a review gate. We'll walk through each phase in detail — with a real example. WALKING THROUGH IT SDD in action — the four phases. 07 / PHASES 1 /specify Specify Write intent in plain prose. AI drafts; you approve. HUMAN + AI 2 /plan Plan AI derives architecture from spec and constraints. HUMAN + AI 3 /tasks Tasks Atomic units, ordered by dependency. AI 4 /implement Implement Agent executes, self- validates, checks off tasks. AI Spec Driven Development The End Of Vibe Coding 9 / 27
10. DETAILS SURFACE CLI · any editor OPINION Low — bring your agent SWEET SPOT New projects, new features DETAILS SURFACE CLI · any editor OPINION Medium — spec deltas track change SWEET SPOT Existing codebases (brownfield) DETAILS SURFACE Dedicated IDE OPINION High — full workflow SWEET SPOT Regulated · large teams PICK YOUR STACK SDD tooling landscape — three approaches, one goal. 08 / TOOLING 01 · CLI · OSS · 88K★ GitHub Spec Kit Open-source CLI making specs executable artifacts. Agent- agnostic (Claude, Copilot, Cursor, Codex). Constitution → Specify → Plan → Tasks → Implement BEST FOR GREENFIELD 02 · FRAMEWORK · 40K★ OpenSpec Universal planning layer for 40+ AI coding tools. Zero infrastructure — just Markdown in your repo. Propose → Apply → Archive BEST FOR BROWNFIELD 03 · IDE · AWS Kiro (AWS) Full IDE with SDD built into the editor. Turnkey workflow, enterprise guardrails. Requirements → Design → Tasks BEST FOR ENTERPRISE Spec Driven Development The End Of Vibe Coding 10 / 27
11. Everything lives in version control. No hidden state, no external SaaS — the repo is the system. SETUP Your repo, your rules — three files and you're running. 09 / SETUP 01 · PROJECT Constitution Project-level rules: tech stack, conventions, boundaries. Every agent reads it before touching code. ./CONSTITUTION.md 02 · PER-FEATURE Specs directory One Markdown file per feature. Plan, tasks, and generated artifacts live next to each spec. .specs/<feature>/spec.md 03 · AGENT Agent config Points the agent at the constitution and spec workflow. Pick the file your tool reads. CLAUDE.md · .cursorrules · AGENTS.md Spec Driven Development The End Of Vibe Coding 11 / 27
12. DESIGN & DEVELOPMENT Prove it works first — write acceptance tests before the feature. Speak the same language — business terms in code, specs, conversation. Contracts before UI — agree on the interface first, build the screen second. One thing, one responsibility — easy to change, hard to break. ARCHITECTURE & QUALITY One team owns one feature — end-to-end, no hand-offs. Security is a default — not a feature request, baked in from day one. Explicit "done" criteria — each phase has a gate, no shortcuts. Track your building blocks — every dependency pinned and auditable. MANDATORY WORKFLOW Specify → Contract → Plan → Tasks → Red → Green → Refactor → Review CONSTITUTION Boundaries before code — the repo constitution. 10 / CONSTITUTION Spec Driven Development The End Of Vibe Coding 12 / 27
13. A spec is a structured Markdown document that captures the three pillars every AI agent needs to succeed — stripping these down causes silent drift. ANATOMY What makes a good specification? 11 / ANATOMY PILLAR 01 Intent = "What are we building and why?" The goal, the user story, and the motivation behind the feature. # Feature · User story # Motivation PILLAR 02 Constraints = "What rules must we follow?" Tech choices, patterns, and boundaries that prevent the agent from drifting off course. ## Tech stack ## Must / Must-not PILLAR 03 Acceptance criteria = "How do we know it's done?" Concrete, testable conditions — the agent checks its own work against these. GIVEN / WHEN / THEN FR-001 · FR-002 · ... Spec Driven Development The End Of Vibe Coding 13 / 27
14. DO Be specific about behavior and edge cases. Define acceptance criteria as testable scenarios. List constraints — especially what not to do. Reference existing patterns in the project. Include error handling expectations. DON'T Describe implementation — let the plan phase derive it. Write vague requirements like "make it fast". Assume the agent knows your codebase or history. Mix multiple features in one spec. Skip non-functional requirements (limits, audit). The spec is the contract. If it's ambiguous, the agent will guess — and usually guess wrong. CRAFT Writing good specs — do's & don'ts. 12 / CRAFT Spec Driven Development The End Of Vibe Coding 14 / 27
15. The Spec You write intent — AI drafts the spec, flags ambiguity FR-004: Evaluation latency MUST NOT exceed [NEEDS CLARIFICATION: target p99 latency?] 3 13 / SPEC [NEEDS CLARIFICATION] AI flags what it cannot assume — no silent guessing 1 Input 2 Draft Flag Spec Driven Development The End Of Vibe Coding 15 / 27
16. The Plan You steer tech direction — AI derives the architecture research.md — tech choices, trade-offs data-model.md — entities & relations contracts/ — OpenAPI spec quickstart.md — dev setup REVIEW GATE You approve plan + artifacts before tasks are generated. 3 14 / PLAN SUPPORTING ARTIFACTS + REVIEW GATE GENERATED ALONGSIDE PLAN.MD You are the gatekeeper at every phase 1 Input 2 Plan Gate Spec Driven Development The End Of Vibe Coding 16 / 27
17. The Tasks AI breaks the plan into atomic, verifiable units T01 Database Schema & Migrations FILE migrations/001_flags.sql DEPENDS — ACCEPTANCE Schema matches data-model.md, audit table append-only VALIDATE npm run migrate && npm run test:schema T02 [P] FlagRepository FILE src/repositories/flag.repository.ts DEPENDS T01 ACCEPTANCE Contract tests pass, archival works, no hard delete VALIDATE npm run test:repo T03 [P] Rollout Evaluator FILE src/services/rollout.service.ts DEPENDS — ACCEPTANCE Hash distribution within ±3% of target over 10k samples VALIDATE npm run test:rollout T04 REST Controller (OpenAPI-bound) FILE src/controllers/flag.controller.ts DEPENDS T02, T03 ACCEPTANCE All endpoints match contracts/openapi.yaml VALIDATE npm run test:contract && npm run test:e2e 2 15 / TASKS SAMPLE TASKS 1 Run Tasks [P] = parallelizable. Agent runs independent tasks concurrently. Spec Driven Development The End Of Vibe Coding 17 / 27
18. Implementation /implement — AI executes, validates, retries 12 / 12 Agent respects dependencies, parallelizes independent tasks, and checks off tasks.md as it progresses. 47 / 47 Self-validation after each task. On failure the agent revisits the task — it does not skip ahead. ~35 min From approved spec to a running service with audit log, cache, and rollout logic. Human review time not included. 0 lines of code hand-written. 100% reviewable trail: spec → plan → tasks → commits. 16 / IMPLEMENT ⌗ TASKS COMPLETED ✓ TESTS GREEN ◷ END-TO-END Spec Driven Development The End Of Vibe Coding 18 / 27
19. Whoever writes the spec is now the programmer. SEAN GROVE "THE NEW CODE" · AI ENGINEER WORLD'S FAIR 2025 Sources: AWS Kiro · AWS re:Invent (2025) · The New Code — AI Engineer (2025) PRACTICAL IMPLEMENTATION Real results. Real adoption. Already. 17 / EVIDENCE " 250k Developers onboarded in 3 months AWS Kiro · The New Stack 94 % Satisfaction across spec-first workflows Delta Airlines · AWS re:Invent (2025) 88k+ GitHub Spec Kit stars github/spec-kit (Apr 2026) Spec Driven Development The End Of Vibe Coding 19 / 27
20. Recap — what just happened Spec → Plan → Tasks → Code YOU + AI Spec ~40 lines of Markdown Your intent, AI-drafted, you reviewed → YOU + AI Plan ~120 lines + artifacts Your tech direction, AI-derived → AI Tasks 12 atomic units AI-generated from plan → AI Code ~1,800 LOC + tests AI-implemented, self- validated Your job shifts from writing code to shaping intent and reviewing artifacts at every gate. 18 / RECAP Spec Driven Development The End Of Vibe Coding 20 / 27
21. A spec is to software what a blueprint is to a building. You wouldn't let the contractor decide the floor plan based on a hallway conversation. Yet that's exactly how most teams use AI coding agents today. WRITE THE BLUEPRINT. THEN HAND OVER THE TOOLS. ≡ Spec Driven Development The End Of Vibe Coding 21 / 27
22. RULE OF THUMB If it takes two or more prompts to explain — write a spec. SWEET SPOTS Where SDD excels. 19 / WHEN TO USE Greenfield projects Spec becomes single source of truth from day one — you design the contract, the agent implements it. Complex features Agents need structure to avoid breaking contracts across services, schemas, and domains. Multi-agent coordination Shared specs prevent conflicting implementations — the spec is the coordination primitive. Compliance & audit "Show me the spec, the review, the commit" — a one-minute answer, not a one-day dig. Spec Driven Development The End Of Vibe Coding 22 / 27
23. RULE OF THUMB If you can explain the task in one sentence — skip the spec. KNOW THE LIMITS When not to use SDD. 20 / WHEN TO SKIP Small bug fixes Spec overhead exceeds the change. A one-line null-check does not need a three-phase review gate. Rapid prototyping Uncertain requirements. Specs need clarity — when you're still exploring, write code, delete it, repeat. UI-heavy / creative Visual design needs iteration, not specs. Pixel choices live in a Figma round-trip, not a Markdown file. One-off scripts Throwaway code — just prompt it. No process needed for disposable scripts. Spec Driven Development The End Of Vibe Coding 23 / 27
24. You don't need to change your whole process. Start with one feature, one spec, one review gate. ACTION PLAN Getting started — Monday morning. 21 / ACTION PLAN DEVELOPER Write one spec for your next feature Pick Spec Kit, OpenSpec, or Kiro. Run the init command. Write a 30-line spec for a feature you'd normally start coding immediately. See the difference. PRODUCT MANAGER Write the "intent" section Describe what should be built and why — in plain prose. Hand the intent to a developer and let the AI draft the spec from there. You review, not reverse-engineer. QA Review acceptance criteria at spec time Don't wait for the PR. Review the GIVEN/WHEN/THEN scenarios in the spec — you'll catch missing edge cases before a single line is written. MANAGER Add review gates to your process Require a spec review before planning, and a plan review before coding. Two gates that prevent weeks of rework and give you audit trail for free. Spec Driven Development The End Of Vibe Coding 24 / 27
25. TAKE HOME Five takeaways. 22 / TAKEAWAYS 01 AI agents are powerful but limited. Under 50% success on realistic tasks — and silent failures when they do fail. 02 Intent as truth — not code as truth. Specs drive everything downstream: plan, tasks, code, tests, review. 03 Specifications become executable — through AI. From passive docs to active source of working software. 04 Choose your tool — Spec Kit · OpenSpec · Kiro. Greenfield · Brownfield · Enterprise — pick once, keep the workflow consistent. 05 Industry convergence on structured intent. SDD is becoming standard practice — GitHub, AWS, and the open-source community are all heading here. Spec Driven Development The End Of Vibe Coding 25 / 27
26. Exceptional at pattern completion, not at mind reading. DEN DELIMARSKY GITHUB PRINCIPAL PM " Spec Driven Development The End Of Vibe Coding 26 / 27
27. Q & A · THANK YOU Thanks for your attention. Let's build software that actually does what it says. Daniel Sogl daniel.sogl@thinktecture.com Principal Consultant · Thinktecture AG linktr.ee/daniel_sogl