Overview

How CommsCrew works.

A guided tour of every user journey in the product, with real screenshots from the live build and honest notes about what's not yet implemented.

The marketing landing at commscrew.genmindz.ai/

What this product is

A four-agent AI system for solo corporate communicators and small teams. The agents — Strategist, Writer, Editor, and Analyst — collaborate on a single pipeline: brief → draft → score → save. From there you edit, route to approvers, schedule, and (eventually) publish.

The whole product runs at <$10/month on AWS, deployed via a single EC2 spot instance + Neon serverless Postgres + Caddy auto-TLS. The build methodology is documented in the Builder’s Playbook.

How to read these docs

Each section is a complete user journey, beginning to end, with screenshots inline.
Every section ends with a red Honest limitations box — what doesn’t work, what’s stubbed, what we know is broken.
Open the demo in a separate tab — the credentials are pre-populated. Click around as you read.

The journeys

Getting started →

From landing page to your first crew message in under 90 seconds.

Creating content →

How a brief turns into three brand-voice-scored variants via SSE streaming.

Editing & approving →

Variant selection, send for review, the approval queue.

Scheduling →

Schedule modal, calendar grid, status transitions.

Memory & learning →

Brand voice profile, org facts, exec ghostwriting, edit-signal learning.

Limitations →

Complete list of what's stubbed, broken, or deliberately deferred.

Honest limitations

The whole product is a deliberately scoped showcase. The most important things to know about before you read further:

Multi-agent is mostly UX framing. Under the hood it’s three sequential Claude calls (Strategist → Writer → Editor) with text passing between them. No autonomous agent behavior.
Publishing is demo-only. Demo OAuth, demo post URLs, and seeded engagement metrics work; real LinkedIn / Twitter OAuth + post APIs are not wired.
Billing is stubbed. Stripe is wired in the model layer but no real payment flow.
Semantic memory is real, but the score isn’t a confidence. Retrieval is genuine cosine similarity over embeddings (recall@5 0.87 on a labelled corpus, against 0.28 for the keyword search it replaced). What it still can’t do is abstain: a query with no answer scores 0.54 against 0.47 for irrelevant documents.
Variant body parsing has a known bug. Currently extracts the “Craft Notes” section instead of the actual body (Issue #4 from the lessons-learned deck).

Getting started →