Originally published on LinkedIn on October 24, 2025.
The article examines the relationship between Spec-Driven Development (SDD) and LLM context length management. While SDD helps break down features and plan implementation with AI coding tools, poorly managed specifications can become counterproductive.
Key Challenges Identified
Single-Shot-Prompts Problem
Large feature requests create massive specifications requiring extensive clarification. Tools like GitHub Spec Kit generate many “[NEEDS CLARIFICATION]” items, and the planning phase can exhaust context windows quickly.
Mountain of Markdown
Generated markdown files can become impractically long for human review. When specifications grow too extensive, developers risk auto-updating them without meaningful review—a trap I call “vibe specifying.”
Lost in the Middle Phenomenon
The article references the Stanford research paper on how language models struggle with information buried in long contexts. Oversized specs can suffer from recency and primacy bias even during the specification phase itself.
Core Recommendations
I propose four key principles:
- Start small: Begin with line-or-two feature descriptions; avoid extended “and” chains
- Specs are for humans first: If you cannot reasonably review it, it’s too large
- Slice meaningfully: Decompose using frameworks like INVEST and MoSCoW prioritization
- Context matters: Design specifications considering agent limitations
Additional Insights
- Feature decomposition requires structured thinking that tooling alone cannot solve
- Subagents can help manage context by maintaining isolated environments and enabling parallelization
- Concepts like EARS (Easy Approach to Requirements Syntax) and user story frameworks remain valuable foundations
Conclusion
SDD, when properly scaled and reviewed, provides benefits over undisciplined “vibe coding” approaches to AI-assisted development.