The agent does the asking.
A guided voice interview by phase: vision, features, brief, design, architecture, quality, security. Every answer fills a field. Every field tracks who answered, when, and why.
AI can write the code. SpecGraph hands it the spec: discovery on the codebase you already run, every constraint and acceptance criterion, sealed into a contract it implements verbatim.
Brownfield-first. Discovery reads the system you already run, six phases pin down what ships next, and your agent cites the pack in every PR.
42 modules · 9 owners · 7 debt flags
constraints · acceptance criteria · tokens
8 files · signed · immutable
implements verbatim · cites the spec
Every panel below is a stylised stand-in for an actual surface in SpecGraph 2.0. The choreography is exactly how the app behaves, just slow enough to read.
A guided voice interview by phase: vision, features, brief, design, architecture, quality, security. Every answer fills a field. Every field tracks who answered, when, and why.
Sketch the UI on Lovable, v0, Base44 or Bolt. SpecGraph reads the mockup, extracts tokens and surfaces, and seeds the brief. The prototype becomes evidence. Not a throwaway.
shadcn-shaped UIs in seconds. Drop the URL or a screenshot. SpecGraph extracts tokens, infers surfaces, and pre-fills the brief.
Prototype the feel on Lovable, v0, or Base44. Ship the plan from SpecGraph. The mockup becomes evidence inside the spec, not a throwaway artifact.
Discovery scans an existing codebase in one pass: modules, ownership, stack, tech-debt, compliance gaps. Then seeds a SpecGraph project on top of it. Same six-phase lifecycle, pre-warmed. Ready for scale, ready for the next feature.
Discovery turns a brownfield repo into a SpecGraph project in one pass. Same six-phase lifecycle, just pre-seeded. From here the team adds new featuresunder the same spec, without retro-fitting briefs by hand.
Engineering, design, compliance, finance, operations. Structured surveys, no calendar warfare. Wishes flow into the spec automatically. Typed, attributed, scoped.
Patient-intake portal · 14 clinics · Epic SMART-on-FHIR · HIPAA in scope.
Patients shouldn't repeat themselves at every clinic.
Drop the legacy bridge. It owns 40% of incidents.
Required by HIPAA + state rules in NY and CA.
Field staff and rural clinics rely on tablets, not desktops.
Anything beyond v1.2 needs a fresh purchase order.
Don't promise a 24/7 SLA we can't staff yet.
Conflicts surface with severity. Three unification strategies sit side by side. Stakeholders vote. Audit log writes itself.
FHIR R4 only, drop HL7v2 bridge
Three of our partners only emit HL7v2
Keep HL7v2 → FHIR adapter for partners A, B, C; sunset roadmap stamped to Q2.
Compare side-by-side, pick a strategy, and the spec rewrites itself. Every decision lands in the audit trail.
When the spec is sealed it ships as a BMAD pack: product-brief, prd, architecture, ux-design, epics, qa-test-plan, project-context, CLAUDE.md, plus per-story implementation. Stamped, immutable, signed.
SpecGraph emits project-specific skills: tiny, scoped subagents that already know your PRD sections, your tokens, and your constraints. Loadable by Claude Code, Codex and Copilot. Your agent stops re-deriving context every session.
Bootstraps SMART-on-FHIR SSO flow with PRD-bound scopes and Epic-specific quirks.
Translates HL7v2 messages into FHIR R4 resources for partners A, B, C. Adapter tagged to sunset.
Enforces 7y, immutable, cryptographically chained audit writes for every PHI mutation.
Form-level field validator wired to PRD §2.1. Knows which fields branch by clinic region.
Re-scans the repo when scope changes; emits a fresh discovery.md and updates open-questions.
The pack and its skills are served over MCP to your coding agent. Claude Code, Codex or Copilot execute it verbatim. They cite the spec back in every PR. The contract is the wire.
Agents stream every started task, every commit, every question back into SpecGraph in real time. Spot drift early. Edit a phase. Re-seal the pack. Agents read the diff over MCP and resume. The loop is closed.
The numbers that matter aren't lighthouse scores. They're how long it takes for a sentence in a meeting to become a tested pull request.
Numbers are anonymized aggregates from current pilots: financial services, healthtech, public-sector logistics. We'll send the full benchmark on request.
SpecGraph isn't one project at a time. The platform layer carries your design system, your tech stack, your compliance framework, and your portfolio of features across every project. Each new spec inherits what the org has already decided.
Upload a design system, an approved tech stack, your security framework, and the templates you reuse. Every new project pulls from the same shelf. No copy-paste, no drift, full attribution back to the source entry.
A single board for the whole org. Score by impact, lock priority by quarter. Each lit row converges into a sealed spec, without losing the frame the leadership team agreed on at the offsite.
Brownfield-friendly. Discovery scans the codebase, infers stack and modules, drafts a discovery.md, seeds vision and architecture, and lists the open questions the voice interview should ask first. Same six-phase lifecycle, pre-warmed.
What gets specified and built
SpecGraph is not only for fullstack apps. Any software artefact with stakeholders, constraints, and an AI agent on the other end benefits from a sealed spec, and from a contract that travels with the code.
Six verticals where a half-written brief costs more than an engineer-month. SpecGraph is built for the room where that cost is already counted.
AI coding agents are spectacular. They are also obedient. They build what the spec says to build. SpecGraph is what makes the spec worth obeying. Every department, on the record. Conflicts resolved before code. One sealed pack, served over MCP.
Private beta. We're onboarding two teams a week.