Me equivoqué, NO uses Spec Driven Development

Summary

The presenter argues that Spec-Driven Development (SDD) has been a poor fit for his work with AI coding agents. In his experience, asking an agent to plan an entire feature in detail can produce lengthy specifications, excessive abstractions, invented requirements, and large code changes that are difficult to review. He also criticizes workflows that build horizontal layers—database first, then API, then frontend—because the software may not be usable until late in the process.

He distinguishes between planning and planning everything in advance. Short requirements and user stories can be useful, but he considers detailed technical designs and implementation plans better suited to delegating work among human developers than to handing an entire project to AI. He also notes that tasks being produced in layers may stem from how language models were trained rather than from SDD itself.

Recommended Approach

The presenter’s alternative combines agile practices, vertical slices, and AI pair programming:

• Maintain a small, revisable backlog of user stories or tickets. Priorities can change, and unnecessary stories can be discarded.
• Refine one story at a time before implementation. Clarify its goal, acceptance criteria, and out-of-scope items without prescribing every technical detail.
• Build vertical slices so that each small increment delivers a usable feature across the relevant frontend, backend, and database layers.
• Use AI as the driver and the developer as the navigator. The AI writes code; the developer decides what to build, reviews changes, and guides design and refactoring.
• Work in short, test-guided iterations. Review small changes, use tests to check behavior, and refactor when there is a concrete reason rather than preemptively.
• Protect tests and review changes. The demo uses frequent Git commits so test modifications are easy to spot and changes remain manageable.

The presenter describes this as augmented coding, not “vibe coding”: AI reduces the amount of code the developer must type, but not the need to think, make decisions, and understand the code.

Demonstration and Tools

The video walks through building a simple multi-user to-do application:

1. Use an AI skill to create a minimal backlog, then split stories that combine multiple functions.
2. Refine one story with questions and acceptance criteria while stating what is out of scope.
3. Implement the story incrementally, starting with tests and reviewing each small change.
4. Add structure or refactor only when the code begins to need it—for example, by moving routes and tests into separate files and moving validation into Zod middleware.
5. Test the finished registration flow in the running application.

The demo uses a project skeleton built with Deno, React, and Drizzle, with Claude for planning and coding, Zod for validation, and Neovim and LazyGit for code review and Git commits. The presenter uses a more capable Claude model for backlog and story refinement, then switches to a faster model for implementation.

The video also names SDD-style tools, including GitHub Spec Kit, Kiro, BMAD (which subtitles may render as “Bimad”), and OpenSpec.

People and Sources Mentioned

• Main speaker: The HolaMundo presenter, who identifies himself as Nico during the demonstration.
• Developers cited in the discussion: Kent Beck, Matt Pocock, and Dave Farley.
• AI tools used in the demonstration: Claude, including the Opus and Sonnet models.