Reference

All
Skills

6 SKILL.md files ~/.claude/skills/<name>/SKILL.md open a card to read the full file
At a glance
CommandUse whenWritesUses agents
/sdd-startup [idea]New project, no code yetsteering docs, CLAUDE.md, STATE.mdnone
/sdd-checkupAny unfamiliar or dormant projectonly the fixes you approveExplore (large repos)
/sdd-bootstrapCode exists, no context filessteering docs, CLAUDE.md, STATE.md3× Explore, in parallel
/sdd-feature <name> [description]Before building any featurespecs/<name>/: requirements, design, tasksExplore, for research
/sdd-implement <feature> [and …]The spec is approved and committedcode, one commit per task, ticks in tasks.mdsdd-implementer per task; worktrees for parallel features
/sdd-validate <feature>Every task is tickedsteering docs, STATE.md, CHANGELOG, README, versionsdd-reviewer
Only you can start these skills. Each one sets disable-model-invocation: true, because all of them write files or commit. Claude can suggest a skill, but it won't run one on its own. Arguments after the command, like the feature name, are passed to the skill as $ARGUMENTS, and autocomplete shows what each skill expects.
Each skill
/sdd-startup
New project
You have an idea but no code. Interviews you, then walks through each stack decision.
+
use when: starting from scratch stops if the folder already has code

Asks about purpose first, then works through each stack area in order: language, framework, data layer, async work, hosting, testing, hard constraints. Each area is a multiple-choice question with 2–3 options, one-line trade-offs, and a recommendation listed first. You make every choice. Then it writes product.md → tech.md → structure.md → CLAUDE.md (last) → STATE.md, and adds .claude/worktrees/ to .gitignore.

/sdd-checkup
Project health audit
You've opened an unfamiliar or dormant project and want to know what's missing or stale.
+
reports first one approval per fix upgrades older SDD projects

Checks the context files and CLAUDE.md size, @ imports, STATE.md drift against git, specs STATE.md doesn't mention, tasks missing Files or Check lines, stale CHANGELOG, README and version, whether the agents are installed, and whether worktrees are gitignored. Ranks the fixes Critical → Important → Optional and applies only those you approve. If it finds an AGENTS.md, it keeps it and adds a CLAUDE.md that imports it.

/sdd-bootstrap
Existing project: write the context files
A codebase exists, but there's no CLAUDE.md or steering docs.
+
once per project run /sdd-checkup first 3 Explore agents in parallel

Three Explore agents map the stack, the structure and the conventions at the same time, and each returns a short summary with file paths. Claude asks only about what the code can't tell it, and never invents decisions. tech.md marks each decision inferred or confirmed. CLAUDE.md is written last and stays short, and STATE.md starts at the current commit.

/sdd-feature
Spec a feature
Requirements → design → tasks with checks, with your approval after each phase. No code.
+
use when: any new feature or multi-file change needs CLAUDE.md and tech.md /sdd-feature csv-export let users export reports

The first word is the feature name and the rest is an optional description. Requirements come from an interview and end with numbered, testable acceptance criteria. Design uses Explore agents for research, checks against tech.md, and stops at every stack gap. Tasks each list their files, a check command and the criteria they cover, so a fresh agent can complete one without the conversation. If the spec folder exists, it resumes where you left off.

/sdd-implement
Implement features
Coordinates the build. One agent per task, then a check, a commit and a tick for each. Parallel worktrees for several features.
+
use when: spec approved and committed resumes from the first unticked task /sdd-implement feat-a and feat-b

The main conversation doesn't write the feature code. It plans, waits for your approval, then gives each task to a new sdd-implementer agent. Tasks in a PARALLEL group run at the same time. After each one it re-runs the check, makes sure only the expected files changed, ticks the task and commits [T<n>]. It retries once on failure, then stops and asks you. With several features it compares their files and runs the independent ones in parallel worktrees. It pre-approves the git commands it needs (allowed-tools), so committing doesn't ask permission each time.

/sdd-validate
Validate a finished feature
An independent reviewer gives a pass/fail verdict, then the docs and STATE.md are updated.
+
use when: every task is ticked same session is fine the first step is a hard stop

It starts by loading git status and the recent log into the skill (the !`command` syntax), so it knows the repository's state without being told. sdd-reviewer gets only the feature folder and the commit range, and returns a pass/fail verdict with evidence for each acceptance criterion. Any failure stops the skill. After a pass, it updates the steering docs, STATE.md, CHANGELOG, README and version, and asks you before a version bump.

Install
Use the one-line install on the Setup page. It fetches every skill and agent into the right folders, and you can re-run it to update. If you download a file here instead, rename it to SKILL.md inside a folder named after the skill, as shown in the modal's save as line.