PRD Is Dead: How SPEC Documents Unlock AI-Driven Product Development

Three Modes of Requirements Production

The author contrasts three approaches:

Traditional PRD: Written for human developers. Relies on implicit knowledge, allowing ambiguity that developers fill in through discussion. Flexible but slow — days per PRD plus review cycles.

Vibe Coding: Direct natural-language prompts to AI (popularized by Andrej Karpathy, 2025). Fast for demos but uncontrolled; complex requirements drift, requiring more correction time than manual coding.

Spec Coding (SPEC-driven): Product managers write a SPEC document — a precise, execution-oriented contract for AI. Contains only core logic, business flows, boundary conditions, acceptance criteria. Prototypes are external HTML files linked to the SPEC. Developers feed the SPEC to AI to generate production-ready code without a human-to-AI translation step.

The Interface Shift: From Human to AI

The fundamental change is the consumer of requirements. Historically, developers read PRDs and used experience to infer missing details. Now developers increasingly pass requirements to AI code generators. Since AI does not infer, the input format must change: every boundary, exception, input/output, and error code must be explicit. A valid SPEC leaves zero ambiguity for the AI reader.

SPEC vs. Structured PRD: Core Differences

PRD assumes human imagination: Only "what to do" is written; "what not to do," edge cases, and exception handling are left for developers to fill in during review meetings.

SPEC assumes zero inference: All boundaries, exceptions, constraints, and error codes are locked down. The document appears simpler — stripped of background, explanations, and embedded prototypes — but has higher information density because every sentence is executable instruction.

Prototypes: In PRDs, prototypes are embedded images for human understanding. In SPEC, prototypes are separate HTML files; AI reads the HTML structure directly to reconstruct page logic.

Origins of SPEC

Roots trace to 1980s Design by Contract (preconditions, postconditions, invariants) and 2000s spec-first/TDD practices. The modern explosion came in late 2025 when Amazon AWS, GitHub Copilot, Google DeepMind Codey, and Tencent Cloud independently hit the same wall: raw AI coding had high rework rates. All converged on "write SPEC first, then generate code." GitHub released Spec Kit; Thoughtworks placed Spec-Driven Development on its Technology Radar. Spec Coding is now a standard AI-era workflow.

Real-World Experiment: Five Requirements in One Morning

The author, with a pre-built domain knowledge base, processed five requirements in parallel using five AI chat windows. Each requirement previously took a full day (PRD + prototype) and represented ~32 developer hours. Total AI cost: ~$50 (foreign LLMs). The "write document + draw prototype" steps were eliminated. Freed time redirected to evaluating whether each requirement should exist, seeking better solutions, and assessing business value.

PM Value Shifts from Expression to Judgment

Expression layer (writing clear docs, drawing prototypes, structuring flows) is now automated by AI.

Judgment layer remains human: defining the real problem, identifying user pain points, deciding what not to build, setting boundaries, designing low-cost validation experiments.

Two essential traits for AI-era PMs: Curiosity (adopting new tools early creates order-of-magnitude gaps) and Professionalism (deep business understanding determines judgment quality).

Caveats and Next Steps

SPEC is not a silver bullet. It requires supporting infrastructure: a personal domain knowledge base so AI-generated SPECs stay grounded, skill configurations, and MCP (Model Context Protocol) connections for context. The full product R&D chain — review, testing, release — must be redesigned around "AI as executor." The author plans follow-up articles on knowledge-base construction and end-to-end AI-era collaboration workflows.