Build This Now
Build This Now
Builds ReaisState of Claude Code 2026: What 2,500 Public Repos RevealConstruir Já Não É o GargaloDistribuição É o Novo MoatPorque o QA é o Verdadeiro Gargalo no Desenvolvimento com IAPrimeiros Princípios na Era dos MVPs de 24 HorasA Curva de Autonomia: Quanta Liberdade Podes Dar a um Agente de IA?Da Ideia ao SaaSGAN LoopHooks Auto-EvolutivosDo Trace à SkillAgentes de DistribuiçãoAgentes de Segurança com IAEnxame Autônomo de IASequências de Email com IAA IA Limpa-se a Si PrópriaAgent Swarm OrchestrationConstrói uma App Completa com Claude Code: Exemplos ReaisClaude Code para Não-Programadores: Exemplos ReaisClaude Code for Freelancers: Ship 3x FasterA Security Update from Build This NowThe AI Agent That Deleted a Production Database in 9 SecondsHow to Build Your Own Claude Code Harness (or Buy One)Run Claude Code on a Cheaper Model: DeepSeek and GLM Cost ArbitrageIs Claude Code Just a Thin Wrapper? Inside the Harness DebateHow Much Does It Really Cost to Build a SaaS with Claude Code?How to Cut Your Claude Code Token Bill in HalfDo I Still Need a Boilerplate If I Use Claude Code?Harness vs Boilerplate vs Framework: The Build-System Stack ExplainedHow Long Does Idea to Production Actually Take with Claude Code?Is Vibe Coding Safe? What the Lovable and Moltbook Breaches TeachOwn Your Vercel Analytics: I Built a Drain-to-Postgres PipelineSpec-Driven Development Explained: Why Pros Stopped Vibe CodingState of Vibe-Coded SaaS Security (2026 Data)From Vibe Coding to Production: The Checklist That Stops Data LeaksVibe Coding vs Vibe Engineering vs Agentic Engineering: The 2026 GlossaryWhat Is an Agent Harness? Why the Harness, Not the Model, Is the 2026 Moat
speedy_devvkoen_salo
Blog/Real Builds/Spec-Driven Development Explained: Why Pros Stopped Vibe Coding

Spec-Driven Development Explained: Why Pros Stopped Vibe Coding

Spec-driven development means writing structured Markdown specs before AI writes code. Here is how GitHub Spec Kit and Kiro made it the 2026 default.

Pare de configurar. Comece a construir.

Templates SaaS com orquestração de IA.

Published Jun 25, 20268 min readReal Builds hub

Spec-driven development is a workflow where you write your requirements, design choices, and a task list as structured Markdown files before any code runs. AI agents then build against those files step by step instead of guessing from a loose chat prompt. In 2026, spec-driven development became the default way pros ship production software, replacing "vibe coding" (typing prompts and accepting whatever the AI returns).

Pare de configurar. Comece a construir.

Templates SaaS com orquestração de IA.

The moment specs went mainstream

Two events in May 2026 marked the turn. GitHub Spec Kit, a free open-source toolkit for spec-driven development, crossed 90,000 stars on GitHub, roughly doubling in six months (GitHub repo stars). At nearly the same time, Amazon moved its spec-native Kiro IDE to broad global availability on May 7, 2026 (AWS announcement). An IDE is the program developers write code in. "Spec-native" means the tool is built around specs first, not chat first.

When two large players ship spec tooling six weeks apart, it stops being a niche habit. It becomes the standard way teams work.

What vibe coding is, and why it felt great

The term "vibe coding" comes from a February 2, 2025 post by AI researcher Andrej Karpathy on X. Collins later named it a Word of the Year. The idea: describe what you want in plain English, let the AI write the code, and ride the vibes.

It worked at first, and the data backed it. A randomized MIT trial found AI coding tools helped developers complete 26% more tasks and produce 38% more successful builds (MIT RCT, reported). For a quick prototype or a throwaway script, vibe coding is genuinely faster.

Then the three-month wall hits.

Why vibe coding breaks at scale

The problems show up once a codebase grows past a weekend project. The numbers are blunt:

  • Teams saw 41% more bugs after adopting AI coding tools (industry analysis, reported).
  • 91.5% of vibe-coded apps studied contained AI-traceable security holes, and 45% failed the OWASP Top 10, a standard checklist of the ten most common web security risks (security study, reported).
  • A METR study found a 40-point gap between how productive developers felt (+20%) and how productive they actually were (-19%) when using AI on large, existing codebases (METR, reported).

That last gap is the trap. Vibe coding feels fast while it quietly creates work. By day 90, fixing those bugs can eat 20% to 30% of a team's sprint capacity (a sprint is a fixed work cycle, usually two weeks).

How spec-driven development actually works

Spec-driven development replaces the guesswork with written contracts. GitHub Spec Kit organizes this into four phases, each producing a Markdown file the AI reads:

  1. Specify. Write what the feature does from the user's point of view. The "what," not the "how."
  2. Plan. Write the architecture: which files, which database tables, which APIs. The "how."
  3. Tasks. Break the plan into a dependency-ordered checklist, so task two never starts before task one it depends on.
  4. Implement. The AI agent executes the tasks one by one. You review each step.

Spec Kit also uses a Constitution file: a short document that locks in non-negotiable rules for the whole project, like "every database table needs row-level security" or "every feature ships with tests." Row-level security is a database rule that stops one user from reading another user's rows. The Constitution travels with every spec, so the AI cannot quietly drop your standards.

This is the same idea behind a CLAUDE.md file that gives an agent persistent project context, and behind Claude Code subagents that each own one job.

Vibe coding vs. spec-driven development

MetricVibe CodingSpec-Driven Development
WorkflowFreeform prompts in a chat boxStructured Markdown artifacts
Code quality41% more bugs reported80-150% speed gain with quality gates
Security91.5% had AI-traceable vulnerabilitiesAuditable, task-by-task execution
ScalabilityThree-month wall, 20-30% of sprint lost to bug fixesSpecs compound as the codebase grows
ToolingAny chat UIGitHub Spec Kit, Kiro, Build This Now
Best forPrototypes and throwaway scriptsProduction SaaS and team codebases

Does it actually ship faster?

Yes, once the project is real. Teams pairing structured specs with AI execution reported a 150% speed gain over a no-AI baseline and an 80% gain over freeform-prompt AI (reported). Kiro case studies describe an 18-month rearchitecture compressed to 76 days using one-fifth the engineers (Kiro case study, reported).

The tradeoff is honest: specs add 20 to 40 minutes of writing before the first line of code. For a one-hour script, skip it. For anything you plan to maintain, that upfront time pays back fast.

Tools you can use today

  • GitHub Spec Kit. Free, MIT-licensed, with 29+ agent integrations including Claude Code. Best if you want the raw method with no lock-in.
  • Kiro IDE. A VS Code-based editor built around specs. Free tier with 50 credits per month and Claude Sonnet 4.5; paid plans start at $20/month.
  • Build This Now Code Kit. The $29 Code Kit is an opinionated harness for Claude Code with the same three-phase arc baked in: /discover writes the product spec, /mvp-spec turns it into a task plan, and /mvp-build executes. It ships with a production SaaS skeleton (auth, Stripe payments, PostgreSQL with row-level security) so the specs build on a real foundation, not an empty folder. It pairs naturally with MCP servers for connecting outside data.

The risk for founders, not just developers

If you run a product, the security number is the one that should stick: a 91.5% vulnerability rate is a liability waiting to surface, often after launch. The three-month wall is a budget line, quietly draining 20-30% of every sprint into rework. Spec-driven development is the difference between a product that gains value as it grows and one that needs a full rewrite by month four.

Vibe coding is a fine way to find out if an idea is worth building. Spec-driven development is how you build the thing once you know it is.

FAQ

What is spec-driven development?

Spec-driven development is a workflow where you write structured requirement, design, and task files in Markdown before running any code. AI agents then execute against those files step by step instead of generating code from vague prompts. The result is code you can audit and maintain.

Vibe coding vs spec-driven development, which is better?

Vibe coding is faster for throwaway prototypes but builds technical debt quickly, with studies showing 20-30% of sprint capacity shifting to bug fixes by day 90. Spec-driven development takes slightly longer upfront but produces auditable, maintainable code. Teams using it report 80-150% speed gains over freeform-prompt AI approaches.

How do I use GitHub Spec Kit with Claude Code?

Install GitHub Spec Kit and run specify init my-project --integration claude-code. It scaffolds the right command files, context rules, and directory structure for Claude Code automatically. You then move through four phases: Specify, Plan, Tasks, and Implement.

Is Kiro IDE free?

Yes. Kiro has a free tier with 50 credits per month and access to Claude Sonnet 4.5. Paid plans start at $20/month (Pro, 1,000 credits). AWS Startups members can apply for up to one year of free Kiro Pro+.

More in Real Builds

  • A IA Limpa-se a Si Própria
    Três workflows noturnos do Claude Code que limpam a própria bagunça da IA: o slop-cleaner remove código morto, o /heal repara branches partidas, o /drift deteta deriva de padrões.
  • Agent Swarm Orchestration
    Four infrastructure layers that stop agent swarms from double-claiming tasks, drifting on field names, and collapsing under merge chaos.
  • GAN Loop
    Um agente gera, outro destrói, e repetem até a pontuação parar de melhorar. Implementação do GAN Loop com definições de agente e templates de rubrica.
  • A Curva de Autonomia: Quanta Liberdade Podes Dar a um Agente de IA?
    A autonomia que podes dar a um agente de IA depende de uma só coisa: quanto tempo um modelo segura uma tarefa sem se desviar. Uma boa estrutura mais um modelo de confiança é o que liberta o trabalho real de agentes.
  • The AI Agent That Deleted a Production Database in 9 Seconds
    An AI deleted PocketOS's production database and all backups in 9 seconds. Here is why it happened and the guardrails that prevent it.
  • Sequências de Email com IA
    Um comando do Claude Code constrói 17 emails de ciclo de vida em 6 sequências, liga gatilhos comportamentais do Inngest e lança um funil de email com ramificações pronto a implementar.

Pare de configurar. Comece a construir.

Templates SaaS com orquestração de IA.

Own Your Vercel Analytics: I Built a Drain-to-Postgres Pipeline

Vercel has no API to query Web Analytics history. The only export path is a Drain. Here is the HMAC-signed receiver, the Postgres rollup, and the one-session build that owns the data.

State of Vibe-Coded SaaS Security (2026 Data)

Vibe coding security statistics from 2026 scans: 5,600 apps, 2,000+ vulnerabilities, 45% OWASP failure rate, and the Moltbook breach.

On this page

The moment specs went mainstream
What vibe coding is, and why it felt great
Why vibe coding breaks at scale
How spec-driven development actually works
Vibe coding vs. spec-driven development
Does it actually ship faster?
Tools you can use today
The risk for founders, not just developers
FAQ
What is spec-driven development?
Vibe coding vs spec-driven development, which is better?
How do I use GitHub Spec Kit with Claude Code?
Is Kiro IDE free?

Pare de configurar. Comece a construir.

Templates SaaS com orquestração de IA.