specmaster

Turn a rough idea — or an existing site/app/doc — into a comprehensive, build-ready PRD + spec you can hand straight to an AI coding agent.

🌐 Website & Get Started →

specmaster is a Claude Code skill. It walks a 17-criterion checklist derived from real engineering standards (ISO/IEC/IEEE 29148, BABOK, DAMA-DMBOK, arc42, C4, ISO 9241, McGovern Top Tasks, PuMP), writes requirements in EARS and acceptance criteria in Gherkin, and produces a spec structured so a coding agent can build from it — and verify it — without guessing.

It runs standalone. No other skills required.

---

The problem it solves

Most "PRDs" are vague prose. An AI agent then fills the gaps with assumptions, and you get plausible-but-wrong software. specmaster forces the spec to cover the things that actually get dropped — the workflow, the top user tasks, the UX, the UI states, the data model, the measurable outcomes — and writes every requirement as a testable statement with an executable acceptance criterion, so "done" is observed, not asserted.

Install

Clone straight into your Claude Code skills directory:

git clone https://github.com/dubedad/specmaster ~/.claude/skills/specmaster

Restart Claude Code (or open a new session). That's it — /specmaster is now available.

Use — 30 seconds

1. Run /specmaster
2. Answer one question: new thing [M] or improve an existing thing [I]? (if [I], give the URL or path)
3. Answer a short interview about intent (problem, users, scope, outcomes)
4. It generates the rest, validates it, shows you the PRD, you approve.

You get a folder:

specs/<slug>-<M|I>-<date>/
  PRD.md                   the spec (your deliverable)
  requirements.ears.md     testable requirements, classified, stable IDs
  acceptance/*.feature     Gherkin acceptance, tagged to each requirement
  artifacts/               workflow · top-tasks · journeys · UI states · logic model · architecture+ADRs · data model+ERD
  VALIDATION.md            quality + traceability + anti-vagueness report
  MANIFEST.md              17-criterion coverage checklist

Two lanes

Lane	When	How it works
[M] MODERNIZE	A new thing, or a genuinely new way / new audience	Interviews you for each criterion
[I] IMPROVE	Optimize an existing site / app / document	Captures an evidence-based as-is baseline first, then specs the delta

What makes it different

• Standards-grade, not vibes. Every one of the 17 criteria is sourced to a recognized standard — see source-registry.md.
• Covers what specs usually drop: workflow, top tasks (with success-rate + time targets), UX journeys, UI states, data model, measurable outcomes.
• EARS + Gherkin. Requirements are testable shall statements; acceptance is executable — so a build agent knows exactly when it's done.
• Works standalone, degrades gracefully, and accelerates if you have the optional skills below.

Optional accelerators

specmaster runs fully on its own. If any of these Claude Code skills / tools are installed, it will use them for richer artifacts, otherwise it uses its inline method:

ba-requirements-analysis · bpmn-from-website · bpm-audit · gsd-ui-phase · pump-logic-modeler · dama-modeling · dama-architecture · prd-authoring · plus engine adapters: dbt, LikeC4, bpmn-js MCP, Figma Dev Mode MCP, StrictDoc.

Built on (credits)

The methodology stands on the shoulders of: EARS (Alistair Mavin), Gherkin/BDD (Cucumber), ISO/IEC/IEEE 29148, BABOK (IIBA), DAMA-DMBOK, arc42 & the C4 model (Simon Brown), MADR, ISO 9241-210, Top Tasks (Gerry McGovern), Jobs-to-be-Done, PuMP (Stacey Barr), the WK Kellogg logic model, and the constitution idea from GitHub Spec Kit.

Test it in 5 minutes

1. Install (above), restart Claude Code.
2. Run /specmaster, choose [M], and give it a small real idea (e.g. "a CLI that renames photo files by EXIF date").
3. Answer the short intent interview.
4. Check the generated specs/.../PRD.md — confirm it has EARS requirements, Gherkin acceptance, and covers workflow / UX / data.
5. For the [I] lane: run it again, choose [I], and point it at a small website — confirm it captures an evidence-based as-is baseline before proposing changes.

Status

v1.1 — hardened after the first real field test: added a freshness gate (a validation is only valid if it post-dates the spec it blesses), strengthened the provenance rule (source claims are produced by opening the source and citing it, byte-auditable), made a PROCESS.md run-map a standard output, and required [ASSUMPTION] tags on any unmeasured target. Actively tested; issues and PRs welcome.

License

MIT — see LICENSE.