AI coding agents are fast, but they drift. They lose context mid-session, hallucinate requirements, and produce changes that are difficult to review. OpenSpec is a workflow layer designed to stop that by requiring written plans in markdown before any code is generated.
How It Works
The core rule is simple: no code until the spec exists. Developers or AI agents write requirements and plans as markdown files first. Only once that documentation is in place does code generation proceed.
The Fission-AI project extends this with a more structured approach. When a change arrives, an isolated folder is created for that specific work. Changes are described as delta specs with clear validation steps attached. Nothing gets merged until the documented validation criteria pass.
What the Pattern Solves
The two problems this targets directly are hallucinations and context loss. Both are common failure modes when AI coding assistants like GitHub Copilot, Cursor, or Claude handle larger or multi-step tasks. Grounding the agent in a written spec gives it a stable reference to return to rather than reconstructing intent from conversation history.
Recent Updates
Version 1.13.1 of the tooling hardened security, improved validation warnings, and made archive behavior more predictable. The release covers more than 30 tools in the OpenSpec ecosystem.
For solo developers or small teams relying on AI agents for significant chunks of their output, the spec-first pattern is worth adopting. The overhead of writing a markdown plan is low. The cost of a hallucinated implementation that touches production is not.
