Skip to content

Getting started

Install the skills, prepare a project with sdd:setup and take a first work item through the pipeline.

Terminal window
claude plugin marketplace add georgeracu/sdd-skills
claude plugin install sdd@georgeracu

Recommended for Claude Code. The skills install as /sdd:spec-workflow, /sdd:setup and so on.

Choose one route per agent. Two routes install every skill twice.

In Claude Code 2.1.275 or later, one command also works: /plugin install sdd --marketplace georgeracu/sdd-skills.

The plugin carries a version number, so an installed plugin changes only when you run claude plugin update sdd@georgeracu after a new release. For a manual copy, check out a tag, such as v0.2.0, before you copy the files. While the version is 0.x, a minor version can contain breaking changes. The changelog marks them.

2. Install superpowers for the full pipeline

Section titled “2. Install superpowers for the full pipeline”

The full feature pipeline needs the superpowers plugin, or an equivalent, today. Step 8 uses its subagent-driven-development skill to implement the tasks, and step 12 uses finishing-a-development-branch to open the pull request. Every other step, including all the gates, is in this repository.

Without superpowers, see Run without superpowers below.

Run this once in the root of your project:

/sdd:setup

The skill looks at the project before it changes anything. Then it does three things, and asks for your approval before it writes:

  1. It copies the missing files of the scaffold into project/work-items/: the process documents and the templates the other skills read. It never overwrites a file that exists.
  2. It checks AGENTS.md for your build and test commands, and asks for them if they are missing. The skills read the commands from there instead of assuming a build tool.
  3. It reports which optional inputs it found: openapi.yaml, ARCHITECTURE.md and UBIQUITOUS_LANGUAGE.md. A missing input is not an error.

The setup page has the full skill.

4. Take a first work item through the pipeline

Section titled “4. Take a first work item through the pipeline”

Start a feature with discovery:

/sdd:spec-discovery

Discovery agrees the work item’s name, problem, scope and approach with you, one question at a time, and writes discovery.md into project/work-items/{NN}-{name}/. After you approve it, the agent continues through the pipeline. Each phase stops for your review, and each gate writes its verdict into the same directory:

  1. spec-generator writes requirements.md, then design.md.
  2. spec-design-review checks the design and writes design-review-summary.md. A BLOCKED verdict stops the pipeline until the design is fixed.
  3. spec-generator writes tasks.md, and spec-task-review checks it.
  4. spec-qa-review in plan mode writes test-plan.md.
  5. superpowers implements the tasks.
  6. spec-implementation-audit compares the code with the spec, using the build and test output as evidence.
  7. spec-qa-review in review mode checks the tests against the plan.
  8. spec-maintenance annotates older specs that the new work supersedes or changes.
  9. superpowers opens the pull request.

The spec-workflow page is the reference for the order and for what each step writes. The worked example shows every artefact from one real feature.

For a bug, start with /sdd:spec-generator instead. It writes bugfix.md, design.md and tasks.md, with no gates.

Without superpowers, the pipeline runs from discovery to the test plan, steps 1 to 7. Then:

  1. Implement the tasks in tasks.md yourself, or with any agent, and tick each box when its task is done.
  2. Run /sdd:spec-implementation-audit. It reads the spec documents and the code, however the code was written.
  3. Continue with /sdd:spec-qa-review in review mode and /sdd:spec-maintenance.
  4. Open the pull request yourself.

Every review skill works from the documents in project/work-items/, so you can also run one gate on its own against a spec you wrote by hand.