LegalCAN I VIBECODE
Harvey?
BUILD FIRST
- A non-sensitive or read-only version of the core experience
- A manual review step wherever correctness or trust matters
- A narrow companion tool that removes a real piece of work
LEAVE FOR LATER
- Regulated decisions, payments, or irreversible actions
- Claims of correctness, compliance, or uptime that the MVP cannot support
- Sensitive integrations until security, support, and auditability are ready
RESEARCH NOTES
What this tool actually does.
Harvey is a purpose-built AI platform for legal and professional services teams. Its core tools are: (1) Assistant — a chat/drafting interface that answers questions, summarizes long documents, generates and edits drafts, with every response linked to exact cited sources (supports LexisNexis, EUR-Lex, EDGAR, web, or the firm's own document vault); (2) Vault — secure bulk document storage and analysis, offering one-click workflow agents (e.g., M&A, litigation), review tables that auto-structure extracted data, targeted Q&A over large datasets, and bulk summarization; (3) Workflow Agents — pre-built or custom agents that run end-to-end tasks in parallel or on schedules, accept documents/images/video/audio, and return review-ready cited reports, presentations, or interactive pages; agents include plan preview/approval, human-in-the-loop 'nudges', and full audit logging; (4) Knowledge — rese
RESEARCHED BUILD PROMPT
Give this to your coding agent.
Build a focused prototype of a Harvey-style legal AI workbench (call it something else — do not copy Harvey's name, logo, or UI). Core workflow: a lawyer opens a matter workspace, uploads contracts/discovery documents into a Vault-like store, asks the Assistant questions in plain language and gets answers with inline source citations, runs a pre-built 'M&A due diligence' agent that extracts key terms into a structured review table, schedules a weekly 'conflict check' agent, and must approve each agent's plan before execution and review its cited output before export.
User-facing requirements: (1) chat interface with markdown responses, clickable citations linking to the source document chunk; (2) document upload (PDF/DOCX) with a browsable vault sidebar; (3) review-table view for structured extraction results; (4) agent catalog page with 'run now' and 'schedule weekly' actions; (5) plan-preview/approve modal before any agent executes; (6) simple audit log screen showing agent steps and citations; (7) permission model with lawyer/admin roles.
Technical requirements: document storage (S3-style), chunking + embedding into a vector DB (e.g., pgvector), RAG-based answer generation with source-span citation, agent orchestration supporting parallel tool calls and cron scheduling (e.g., Temporal or a simple APScheduler worker), LLM calls via a pluggable provider interface, and an audit-trail table. Use a pragmatic stack: Next.js + FastAPI + Postgres. Implement mock connectors for external sources (not real LexisNexis/EDGAR).
Be explicit: do not replicate Harvey's proprietary branding, visual design, or exact feature names. Out of scope: real integrations, production security/compliance certifications, billing, multi-tenant SaaS deployment, and any claim of legal accuracy. Deliver a README covering setup steps, architecture overview, and important limitations (hallucination risk, no legal advice, citation reliability caveats, single-user scale).