Clausemaster

Clausemaster turns legal documents into installable AI skills. Upload a contract or a compliance questionnaire, review what the analysis found, and publish a skill that runs in the TODO.LAW suite and in any LegalQuants-community runtime.

Built by TODO.LAW. Skills you author are yours: open format, no lock-in, no revenue cut.

Full documentation lives at clausemaster.todo.law/docs; the End-to-End Guide walks the whole journey from upload to running your skill in todo.law apps and LegalQuants runtimes.

What it does

1. Analyzes your document. A three-step AI pipeline classifies the document (type, jurisdiction, parties), extracts every clause with a plain-language summary and bias assessment, and flags issues: missing clauses, unusual terms, jurisdiction concerns, compliance risks.
2. Generates a skill draft. For contracts, it builds a clause library where each clause carries 3 to 5 negotiation options with pros, cons, and full legal text. For compliance documents (DPIAs, vendor audits, checklists), it builds a structured assessment template with scoring guidance. Every draft also gets evaluation cases grounded in its own content, so quality can be checked before anyone relies on it.
3. Publishes a dual-format skill. One export produces both formats:
   • the todo.law engine files (clauses.json, boilerplate.json or template.json, manifest.json), installable as a signed package in Dealroom or DPO Central (the AI Sentinel installer is planned)
   • an agent skill (SKILL.md with YAML frontmatter, README.md, parameters.json, evals/), loadable by LQ.AI and other LegalQuants-community runtimes

A conformance check runs before every publish, so a malformed skill fails at export time with a clear message instead of failing at install time.

Compatibility

Target	Status	Notes
Dealroom (todo.law)	Yes	Signed .skill package, installer-verified
DPO Central (todo.law)	Yes	Assessment templates with type routing
AI Sentinel (todo.law)	Planned	Installer in development
LQ.AI / LegalQuants runtimes	Yes	Loads as a standard skill folder
Claude and other agent-skill hosts	Yes	SKILL.md follows the agent-skills convention (top-level name and description); extra JSON files are ignored by hosts that do not use them

The skill format is an open specification. Anyone can implement it, and skills are separate works from any host that loads them.

How it compares to conversational skill builders

Anthropic's skill-creator (and the builders based on it, including those used by legal skill hubs) authors a skill through conversation: you describe your workflow, answer questions one at a time, and the agent writes the skill from what you articulate. That works well when the knowledge lives in your head and nowhere else.

Clausemaster starts from the other end: the knowledge is already in your documents. A clause library refined over fifty deals cannot be dictated in a chat session, but it can be uploaded. The two approaches are complementary.

• Use Clausemaster when you have precedent documents, templates, or completed questionnaires. You get structured clause options, an assessment engine, and dual-format output in minutes.
• Use a conversational builder when you are capturing judgment and process rather than documents: escalation rules, negotiation discipline, review workflows.
• Combine them: generate the skill from your document here, then refine the SKILL.md prose conversationally in any agent that supports skills.

Getting the most out of it

• Feed it your best precedent, not a blank template. The clause options and bias analysis are only as good as the source document.
• Review the draft before publishing. Every draft stops in a review state where you can edit clauses, boilerplate, and metadata.
• Check the generated evals. They cite the skill's own clauses and criteria with verifiable assertions, but they are AI-drafted. Have a lawyer confirm any legal claims before relying on them.
• Pick the right AI tier. The free community tier uses a platform-provided open model. Bring your own API key (Anthropic, OpenAI, Groq, Mistral, or Together) for stronger models; your key is encrypted and your documents go to your provider, not ours.

Tech stack

Layer	Technology
Framework	Next.js 16, TypeScript, React 19
API	tRPC 11 with superjson
Database	PostgreSQL via Prisma
Auth	NextAuth 4 (magic link + Google OAuth)
AI	Vercel AI SDK (@ai-sdk/anthropic, @ai-sdk/openai-compatible)
Styling	Tailwind CSS 4, shadcn/ui
i18n	next-intl (EN/ES)
File parsing	pdf-parse, mammoth

Getting started

Prerequisites: Node.js 20+, a PostgreSQL database, and an AI provider key.

git clone https://github.com/RINDOGATAN/clausemaster.git
cd clausemaster
npm install
cp .env.example .env   # fill in DATABASE_URL, NEXTAUTH_SECRET, and an AI key
npx prisma db push
npm run dev            # http://localhost:3002

All AI pipelines run as client-driven steps: the browser calls a "run one step" endpoint in a loop until the server reports done. Steps are idempotent, so an interrupted pipeline resumes when the page reloads, and every AI call fits in a single short serverless invocation.

Useful checks:

npx tsc --noEmit                                  # typecheck
node scripts/check-skill-conformance.mjs <dir>    # validate an exported skill
curl localhost:3002/api/health                    # DB-touching health check

Project structure

src/
  app/                  # Next.js app router
    (auth)/             # Sign-in, verify-request, auth-error
    (dashboard)/        # Documents, upload, analysis, skill drafts
    api/                # NextAuth, tRPC, upload, health
  components/           # Analysis, documents, upload, landing, shadcn/ui
  server/
    routers/            # tRPC routers (document, analysis, skill-draft)
    services/
      ai/               # Provider factory, analyzer, generators, schemas
      document/         # PDF/DOCX/TXT text extraction
      knowledge/        # Reference skill loader
      export/           # Skill exporter and conformance validation
  messages/             # i18n strings (en.json, es.json)
scripts/
  check-skill-conformance.mjs   # Static validator for exported skills

Deployment

Runs on Vercel at clausemaster.todo.law. The filesystem is read-only there, so uploads are stored as bytes in the database, and publishing commits skill files to a Git repository through the GitHub API.

License

AGPL-3.0-or-later. Skills you export are licensed separately, under the license you choose.