OpenSpec Makes You Agree on What to Build Before the Agent Writes a Line
Anyone who's let a coding agent loose on a real feature knows the failure mode: it confidently builds the wrong thing, fast. OpenSpec, trending today at 57,000 stars, is a clean answer. It adds a lightweight spec layer so you and the agent agree on what to build before a single line of code gets written. Each change gets its own folder holding a proposal, the specs, the design and the task list, and the agent works against that instead of against your one vague sentence.
The 1.0 release moved it from experiment to stable and rebuilt the workflow around an action-based system, meaning the tool understands your project's current state: what artifacts exist, what's ready to create next, what each step unlocks. There's a propose workflow that turns a single request into a full change proposal with design, specs and tasks, and a profiles system so you can run just the four essential workflows or a custom set. No rigid phase gates, you can revise any artifact at any time.
The reach is the tell. OpenSpec supports 21-plus tools, Claude Code, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Amazon Q, Cline, even AWS Kiro and the Pi agent, generating the right prompts and skills for each. It's deliberately not tied to one agent. It's a discipline layer that sits above whichever agent you happen to use.
This is the same realization showing up everywhere right now: the bottleneck on agent coding isn't the model, it's the spec. Google Labs shipped DESIGN.md for visual identity, the AGENTS.md and CLAUDE.md conventions standardized project context, and now OpenSpec standardizes the change itself. The pattern is consistent and a little humbling. As the agents get better at writing code, the scarce skill becomes knowing exactly what you want, written down clearly enough that a machine can't misread it.
Link: https://github.com/Fission-AI/OpenSpec
← Back to all articles
The 1.0 release moved it from experiment to stable and rebuilt the workflow around an action-based system, meaning the tool understands your project's current state: what artifacts exist, what's ready to create next, what each step unlocks. There's a propose workflow that turns a single request into a full change proposal with design, specs and tasks, and a profiles system so you can run just the four essential workflows or a custom set. No rigid phase gates, you can revise any artifact at any time.
The reach is the tell. OpenSpec supports 21-plus tools, Claude Code, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Amazon Q, Cline, even AWS Kiro and the Pi agent, generating the right prompts and skills for each. It's deliberately not tied to one agent. It's a discipline layer that sits above whichever agent you happen to use.
This is the same realization showing up everywhere right now: the bottleneck on agent coding isn't the model, it's the spec. Google Labs shipped DESIGN.md for visual identity, the AGENTS.md and CLAUDE.md conventions standardized project context, and now OpenSpec standardizes the change itself. The pattern is consistent and a little humbling. As the agents get better at writing code, the scarce skill becomes knowing exactly what you want, written down clearly enough that a machine can't misread it.
Link: https://github.com/Fission-AI/OpenSpec
Comments