Specdd is a new agent skill that turns feature requests into specs before code is written, according to its GitHub project page. It is designed to work with Claude, Codex, Cursor, Copilot, Windsurf, Gemini CLI and other agents that read SKILL.md or AGENTS.md-style instructions.
Specdd is an open source skill that adds spec-driven development to AI coding agents, according to its GitHub project page. The project says it is meant to turn vague feature requests into production-grade specs before any code is written.
The tool is aimed at people using modern coding agents such as Claude, Codex, Copilot, Cursor, Windsurf, and Gemini CLI. Specdd’s author says it works with any system that reads SKILL.md or AGENTS.md-style instructions.
Specdd is part of a broader shift toward making AI coding agents ask better questions before they write code. A “spec” is a written description of what software should do, while “spec-driven development” means defining that behavior first and building from it. For non-experts, the main idea is that the tool tries to reduce bad AI-generated code by forcing clarification early.
If you use self-hosted agents, add a pre-build spec step to your workflow so prompts are not sent straight to code generation. Copy specdd’s pattern: triage the request, ask for missing edge cases, then require a signed-off spec before the agent can edit files. For automation projects, make your checklist explicit about empty states, retries, error handling, and observability so those checks are part of the agent’s definition of done.
The workflow starts with triage. When a user asks for a feature, specdd first decides whether the request is a skip, quick task, feature, or project, so the scope is visible early and can be corrected.
Next comes an interview phase. Instead of guessing at missing details, the skill asks two to five questions it thinks are essential, such as edge cases, success states, or other requirements that the user did not spell out.
After that, specdd produces a tech-free spec, according to the project page. That spec focuses on what the feature should do and why it matters for the user, rather than how it will be implemented.
The next step is a short technical plan. The project says this plan explains the approach, files, decisions, rationale, alternatives, and risks, and is meant to take less than two minutes to read.
Only after the user approves the spec and plan does the agent build the feature. Specdd then runs a production-grade checklist before declaring the task done, including checks for empty, error, loading, and mobile states, accessibility, keyboard handling, focus management, and observability.
The project also bakes in what it calls UX defaults, including a five-states rule, confirmation patterns, optimistic updates, and accessible focus management. For quick tasks, the skill can keep the output to a short inline mini-spec; for larger work, it expands into separate spec, plan, and task files.
Specdd’s author says the goal is to avoid the two common failure modes of AI-assisted coding. One is “vibe-coding,” where a one-line prompt produces code that looks right but misses basic product and UX details. The other is heavier spec frameworks that add more structure than a small project needs.
The project page says specdd can be installed by copying the skill folder into a supported agent’s skills or instructions directory. It also provides examples for quick features, feature-level work, and multi-day projects, along with references for templates and checklists.
Specdd is published on GitHub under the MIT license. The author says issues and pull requests are welcome, especially for interview playbooks, install instructions for more agents, and examples from real projects.
