Introduction

Spec-driven planning. Parallel TDD execution. A wiki that never forgets.

Most teams using AI coding agents hit the same wall: the agent hallucinates architecture it didn't build, re-solves problems solved three sessions ago, and contradicts decisions made last week. The context window resets. The knowledge vanishes. Every session starts from zero.

SuperSpecs is the discipline layer that makes AI-driven development actually compound.

Three pillars — each solving a distinct failure mode:

Spec

Before any code exists, intent is captured as testable, machine-readable requirements: GIVEN/WHEN/THEN scenarios, SHALL statements, explicit non-goals. The spec fits a fresh 200k context window. Every executor starts with the full contract — not a partial reconstruction of chat history.

TDD

RED before GREEN, without exception. Tests are written first. Code written without a failing test gets deleted. The test suite is the ground truth, not the agent's confidence.

Wiki

Andrej Karpathy's insight, applied at the project level: the context window is ephemeral, but the wiki is permanent. After every shipped feature, knowledge is distilled into structured pages — architecture decisions, patterns, trade-offs, gotchas, open questions. The next session opens the wiki and starts informed, not blind. Knowledge compounds. Problems stay solved. The codebase has memory.

The result: five AI agents running in parallel, each with a fresh context, each working from the same spec, each contributing to the same wiki. Nothing lost between sessions. Nothing re-discovered twice. Every shipped feature leaves the system more legible than it found it.