speccraft

speccraft is a CLI scaffolding tool for spec-driven development (SDD) — Rishabh Software Private Limited's stage-gated workflow that takes one business spec through LLD, decomposition, implementation, testing, and validation. Drop it into any project and pair it with the agent tooling you already use: Claude Code, Gemini CLI, or GitHub Copilot.

Quickstart

npx @rspl/speccraft my-app

What it does

• Drops in spec/ — the full stage-gated SDD workflow (business spec → LLD → decompose → implement → test → validate), agent-agnostic and always installed.
• Ships /speccraft.orchestrate to run that whole workflow end to end from one business spec file. Run it plain for a human-reviewed pass with an approval gate at every stage, or as /speccraft.orchestrate auto <business-spec-file>.md for a fully unattended run — the agent reviews and resolves every gate itself and logs each decision (including how it resolved any conflict) for you to audit afterward. Every stage command it chains through (/speccraft.tech-design, /speccraft.decompose, /speccraft.implement, /speccraft.unit-test, /speccraft.integration-test, /speccraft.validate) also supports the same auto <input> form standalone, so a run never stalls waiting for a human at any gate. See docs/command-reference.md.
• Ships guided setup commands so you never hand-author stack decisions: /speccraft.plan-foundation <business-spec-file> "<tech stack>" drafts the architecture foundation for a new project from a business spec plus a stated tech stack (Layer Scope, locked stack decisions, seed rules/skills, scaffold contract), while /speccraft.onboard [path] does the same for an existing codebase by scanning it instead. Both stop at one human review gate — nothing is locked until you say approved. Run /speccraft.plan-foundation auto <business-spec-file> "<tech stack>" instead to lay the foundation unattended — the agent reviews and writes its own draft directly, resolving an unclear Layer Scope or a contradictory stack description itself and logging how, instead of pausing for you.
• Lets you pick which agent tooling to add: Claude Code, Gemini CLI, GitHub Copilot, and/or a plain agent-agnostic skills layer — install only what you use.
• Selecting Claude Code or Gemini CLI also makes their 13 commands available globally (~/.claude/commands/, ~/.gemini/commands/) in the same run, never overwriting a command you already have.
• Everything is copy-once and yours: no ongoing dependency on this tool after scaffolding, and re-running it never silently clobbers existing files.

Live Examples

Two fully-built, runnable projects showing the workflow end to end — real generated spec/, real application code, every stage gate actually reviewed by a human, not mockups.

| | | |---|---| | examples/greenfield-demo/ | Brand-new project: stack locked from a blank ARCH-DECISIONS.md, one story (a task tracker) built start to finish. See its WALKTHROUGH.md. | | examples/brownfield-demo/ | Existing app onboarded via /speccraft.onboard, then extended with a new story on top of the discovered stack and conventions. See its WALKTHROUGH.md. |

Docs

| | | |---|---| | docs/setup.md | Prerequisites + first-run walkthrough | | docs/greenfield-setup.md | Using the workflow on a new project | | docs/brownfield-setup.md | Using the workflow on an existing project | | docs/usage.md | Every CLI flag, non-interactive usage | | docs/architecture.md | How the tool works, for contributors | | docs/command-reference.md | Every command shipped, per agent | | docs/faq.md | Common questions | | docs/publishing.md | npm publish flow (maintainers) | | troubleshoot/README.md | Known failure modes, symptom → fix | | examples/ | Static install-selection trees plus the two live demos above |

Contributing

See CONTRIBUTING.md. Security issues: see SECURITY.md, not public issues.