Skip to main content

​
Flux Specs

This site is the single source of truth for Flux’s product, architecture, and implementation specs. Every feature, architectural decision, and implementation cycle lives here as native MDX in the employ-inc/flux monorepo. Mintlify renders, indexes, and exposes them — both to humans (in this site) and to AI agents (via the auto-generated MCP server, llms.txt, and contextual buttons).

​
The three-document model

Flux uses three spec types, each with a clear owner and audience:

PRDs

Product Requirements Documents. PM-owned. Captures what to build and why.

RFCs

Requests for Comments. Engineering-owned. Captures how to build it and why this approach.

Cycles

Implementation specs. Engineer + AI authored. Captures the plan to build it.

​
The spec-driven development loop

PRD          (PM)   What to build & why
  ↓
RFC          (Eng)  How to build it & why this approach   (optional)
  ↓
Cycle Doc    (Eng)  Implementation plan, file index, validation
  ↓
Code PR      (Eng + Claude Code)  The actual change
  ↓
Autopilot    (Mintlify) Detects code-spec drift; opens follow-up PRs
  ↓
Cycle Done   Specs marked `completed`; feature shipped
The full process is documented in Cycle 214 — Spec-driven dev with Mintlify.

​
How AI agents use this site

  • Direct file read (authoritative) — Cycle docs, PRDs, and RFCs are MDX/Markdown files in the monorepo. Claude Code, Cursor, and other coding agents read them as native files. Zero translation layer. This is the canonical path.
  • Auto-MCP server (future — once Mintlify site is live and auth model is verified) — Every Mintlify site exposes an MCP endpoint at /mcp. Because the site is gated by Google Workspace auth, MCP/llms.txt access for AI agents requires confirming Mintlify’s API-key or public-carve-out behavior before a flux-specs entry is added to .mcp.json.
  • llms.txt / llms-full.txt — Mintlify auto-generates these at the site root when the project is deployed. Same auth caveat applies until resolved.
  • Contextual buttons — Once the site is live, every page will expose Copy / Open in Claude / Open in Cursor / MCP buttons (configured in docs.json).
  • AGENTS.md at the repo root configures AI agent behavior — both for code agents (Claude Code, Cursor) and for Mintlify Autopilot when enabled.

​
Core conventions

  • PRDs use sequential PRD-NNN numbering — never reused
  • RFCs use sequential RFC-NNN numbering — never reused
  • Cycles use the GitHub issue number as the cycle ID (per Cycle 381) — collision-proof
  • Every Cycle doc references its source PRD(s) and RFC(s) in frontmatter
  • Every PR description references its Cycle number (via Closes #N)
  • Spec drift is a bug — Mintlify Autopilot detects code changes that affect specs and opens follow-up PRs
What I want to doWhere to go
Write a PRDPRDs README → copy _template.mdx
Write an RFCRFCs README → copy _template.mdx
Write a Cycle docCycle template
Browse merged CyclesCycles directory
See what shippedChangelog
Read the AI-Native Manifestodocs/engineering/ai-native-manifesto.md