Harvey Content Fabric

AI-powered LinkedIn content and social engagement tool for B2B sales teams.

Features

• Trend Discovery — daily web search + Reddit scan for trending topics in your niche, filtered by source
• Post Generation — stream multiple LinkedIn posts per trend with tone, size, and humanity-level controls
• Carousel Format — generate 7-slide LinkedIn document posts (hook + 5 content + CTA), avg 6.6% engagement vs 2% for text
• Post Scorer — AI quality analysis on 4 dimensions: hook strength, dwell time, comment magnet, algorithm fit
• Hook Variants — 3 alternative first-line hooks per post (statistic / contrarian / pattern-interrupt)
• CTA Analyzer — flags generic CTAs and suggests specific question-based alternatives for 15+ word replies
• Comment Generator — Reddit and LinkedIn comments using PRD-compliant archetypes; AI highlights best fit
• Reddit Post Finder — search B2B subreddits for high-engagement threads, click to pre-fill comment generator
• Voice Learning — your own saved posts (source="own") inject as voice examples; curated posts inject as structural templates
• Research — summarize URLs and PDFs, write data-driven posts from research notes
• Examples Library — curated post library with AI-extracted metadata (hook type, why it works, engagement tier)
• Knowledge Base — upload PDFs and URLs to inject company context into every generation
• Post History — all generated posts saved with thumbs up/down feedback
• Comment History — all saved comments browsable by platform
• Dual AI Provider — switch between Anthropic Claude (claude-sonnet-4-6) and OpenAI (gpt-4o) in Settings

Stack

Layer	Technology
Framework	Next.js 16 (Turbopack), TypeScript
UI	Tailwind CSS v4, Shadcn UI
AI	Anthropic claude-sonnet-4-6 or OpenAI gpt-4o
Database	Supabase (PostgreSQL), RLS enabled
Deploy	Vercel + cron job for daily trend refresh

Setup

1. Clone and install

git clone https://github.com/labumbu/-linkedin-content-automation
cd linkedin-content-automation
npm install

2. Environment variables

Create .env.local in the project root:

# AI providers — at least one required
ANTHROPIC_API_KEY=          # console.anthropic.com
OPENAI_API_KEY=             # platform.openai.com

# Supabase
NEXT_PUBLIC_SUPABASE_URL=   # your-project.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=

# Cron security
CRON_SECRET=                # any random string

Never run vercel env pull — it overwrites .env.local and wipes local-only keys.

3. Supabase database

Run the schema in your Supabase SQL Editor:

lib/supabase/schema.sql

4. Run locally

npm run dev     # http://localhost:3000
npm run build   # full TypeScript check before pushing

Project Structure

app/
  page.tsx                  # Dashboard — trends with source filter + content accessibility badge
  generate/page.tsx         # Post generator (streaming) — text + carousel formats
  comments/page.tsx         # Reddit + LinkedIn comment generator + Find Threads + history
  research/page.tsx         # Summarize URLs/PDFs + write from research notes
  history/page.tsx          # Saved posts
  settings/page.tsx         # Brand, Topics, Sources, Knowledge Base, Examples, System Prompt

components/
  post-card.tsx             # Generated post — copy, feedback, hook alternatives, CTA note, scorer
  carousel-card.tsx         # Carousel post — slide viewer with nav dots, per-slide copy, caption

lib/
  ai/index.ts               # Provider router — all AI calls go through here
  settings.ts               # Settings helpers + system prompt builder with voice learning
  schemas.ts                # Zod validation schemas
  rate-limit.ts             # In-memory rate limiter
  html.ts                   # Shared HTML stripping utility

Post Generation Formats

Format	Characters	Use case
Short	400–600	Quick punchy takes
Medium	700–1,300	Research sweet spot
Long	1,200–1,600	Full story + data
Carousel	7 slides	Highest engagement format — avg 6.6% vs 2% for text

Voice Learning

The examples library (Settings → Examples) separates your own posts from curated ones:

• source = "own" → injected first as "YOUR OWN VOICE — match this exact writing style"
• source = "curated" → injected after as "STRUCTURAL TEMPLATES — study these patterns"

Up to 8 examples total per generation call. Tone-matching prioritized within each group.

Post Scorer

Click the bar chart icon (📊) on any generated post to get an AI quality score:

Dimension	Max	What it measures
Hook strength	25	First-line impact — specific, pattern-interrupting
Dwell time	25	Readability, white space, reading level
Comment magnet	25	Closing question specificity and reply-chain potential
Algorithm fit	25	Hashtag count, character range, no engagement bait

Cron Job

Trends refresh daily at 06:00 UTC. Configured in vercel.json. The endpoint /api/cron/trends is secured with the CRON_SECRET header.

To trigger manually:

curl -H "Authorization: Bearer YOUR_CRON_SECRET" https://your-app.vercel.app/api/cron/trends

AI Provider Pattern

Every API route that calls AI follows this pattern — never hardcode a provider:

import { resolveProvider, AIProvider } from "@/lib/ai"
const settings = await getSettings()
const provider = resolveProvider(settings?.ai_provider as AIProvider)
// pass provider to lib/ai/index.ts functions

resolveProvider() checks which API keys are set and falls back gracefully.