🎛️ ContentForge

A Cloudflare-native studio for planning, generating, reviewing, and scheduling social content — built as a single full-stack app on Workers, D1, R2, Queues, and Durable Objects.

Brief in, scheduled posts out — with an audit ledger on every pipeline step and a human gate on every outbound send.

At a Glance · Features · Pipeline · Architecture · Getting Started · Scope · FAQ

---

⚡ At a Glance

Question	Answer
What problem does this solve?	Getting from a creative brief to reviewed, scheduled social content usually means stitching together generators, review passes, a media library, and a scheduler. ContentForge runs that whole loop as one serverless app — without reinventing per-platform OAuth.
What is implemented?	The end-to-end brief → concept → prompt → generation → normalize → quality-review → export/schedule loop, a D1 audit ledger with input/output hashes per step, Workers AI safety review, an R2 media library, brand profiles, SEO/competitor research, a compliance-gated sales workflow, planning + Postiz publishing, multi-provider chat, and a live health board.
What is not implemented?	Multi-tenancy (this is a working single-tenant build), public hosting, and native per-platform OAuth (deliberately delegated to Postiz). The video editor and some Intel panels are prototype-grade — labeled honestly throughout.
How do I reproduce it?	wrangler d1 migrations apply contentforge-prod --local && wrangler dev + npm run dev in web/ — no real keys needed to boot.
What production lesson is baked in?	Treat the publisher as a black box behind an API key: the app survives Postiz outages and can swap schedulers by changing a single consumer file.

📑 Table of Contents

• At a Glance
• Overview
• How it compares
• Features
• Tech Stack
• Content Pipeline
• Architecture
  • Worker & primitives
  • The sales approval gate
  • Design decisions
• Getting Started
• Configuration
• Scope & Status
• FAQ
• Glossary
• Contributing
• License

🔎 Overview

ContentForge takes a creative brief and runs it through a multi-stage pipeline: brief intake → concept generation → prompt building → image/video generation → automated quality review → export and scheduling. Everything is stored in Cloudflare primitives so the app scales without servers to manage, and publishing is delegated to Postiz (a self-hostable social scheduler) so we don't reinvent per-platform OAuth.

The demo ships with a sample brand ("Acme", a field-sales/roofing marketing company) to give the studio realistic content to work with. Swap the brand profile, palette, and seed data to point it at anything else — see docs/REPLICATE-FOR-ANOTHER-COMPANY.md.

Note

Status: working single-tenant build. The core generate → review → export → schedule loop runs end-to-end. Some surfaces (video editor, some Intel panels) are prototype-grade. Labeled honestly throughout.

⚖️ How It Compares

	All-in-one social SaaS	ContentForge
Hosting	Their servers, their tenancy	Your Cloudflare account, serverless primitives
Platform OAuth	Built and maintained in-house	Delegated to Postiz behind one API key
Pipeline auditability	Opaque	Every step in a D1 audit ledger with input/output hashes
Generated-content review	Varies	Captioned + safety-classified (LLaVA + Llama Guard) before surfacing
Outbound sales sends	Automated	Human approval required before anything sends

🌐 Live Demo

Not currently hosted publicly. Screenshots below; run locally with the steps under Getting Started.

📸 Screenshots

Login	Command center dashboard

Brand profile — empty state	Brand profile — reading view

More views (wide / mobile login)

✨ Features

• Studio — unified workspace to compose a brief, generate images/video, enhance, and mark posts ready. Live status streams into the generation grid over SSE.
• Content pipeline — a staged brief → concept → prompt-schema → provider-dispatch → normalize → quality-review → export flow. Each step is recorded in a D1 audit ledger with input/output hashes so any run is reproducible and inspectable.
• Automated quality review — generated assets are captioned and safety-classified via Workers AI (LLaVA + Llama Guard) before they're surfaced; low-scoring assets can trigger a delta-prompt regeneration.
• Media library — R2-backed, with drag-drop bulk upload (parallel, worker-proxied) and a reusable picker for reference images.
• Brand profile — voice, palette, products, and forbidden-claim rules that feed every LLM prompt. Cached in KV, embedded in Vectorize for similarity lookups.
• Research & intel — LLM-driven SEO keyword research (with a 24h KV cache) and competitor battlecards backed by Vectorize cosine similarity.
• Sales workflow — a compliance-gated prospect discovery → enrichment → outreach-draft → human-approval → send-queue sequence (public sources only; human approval required before anything sends).
• Planning & scheduling — a D1-backed weekly calendar and a publish queue that hands off to Postiz. A per-minute cron reconciles near-term scheduled posts.
• Multi-provider LLM chat — a context-aware chat slide-over reachable from any tab, routed through Cloudflare AI Gateway (OpenAI / Anthropic / Gemini).
• System status — a live health board that checks D1, R2, KV, Queues, AI Gateway, and Postiz connectivity.

⬆ back to top

🧰 Tech Stack

Layer	Choice
Frontend	React 19, Vite 6, Tailwind CSS v4, Motion, TypeScript
Backend	Cloudflare Workers (Hono router), TypeScript
Data & infra	D1 (SQLite), R2 (object storage), Queues (+ DLQ), Durable Objects (per-user SSE), KV (cache), Vectorize (embeddings), Cron triggers
AI	Cloudflare AI Gateway → Workers AI (gpt-image, chat, LLaVA, Llama Guard, BGE embeddings); Gemini and Replicate as optional providers
Publishing	Postiz public API (self-hosted)
Video tooling	A standalone Python script (tools/generate_videos.py) for image→video generation via Runway / Veo / OpenAI + FFmpeg

🔄 Content Pipeline

%%{init: {"theme": "base", "themeVariables": {"primaryColor": "#0D1622", "primaryTextColor": "#EEF3F7", "primaryBorderColor": "#3B9EF3", "lineColor": "#3078B4"}}}%%
flowchart LR
    B["Brief intake"] --> C["Concept generation"] --> P["Prompt schema"] --> G["Provider dispatch — image / video"] --> N["Normalize"] --> Q["Quality review — LLaVA + Llama Guard"] --> E["Export & schedule"]
    Q -.->|"low score → delta-prompt regeneration"| G

    classDef review fill:#0D1622,stroke:#8EC7FF,stroke-width:3px,color:#8EC7FF
    class Q review

Loading

Each step is recorded in a D1 audit ledger with input/output hashes, so any run is reproducible and inspectable. Nothing reaches the studio surface without passing the review stage — and a low score doesn't just reject an asset, it can feed a delta prompt back into dispatch.

🏗️ Architecture

%%{init: {"theme": "base", "themeVariables": {"primaryColor": "#0D1622", "primaryTextColor": "#EEF3F7", "primaryBorderColor": "#3B9EF3", "lineColor": "#3078B4"}}}%%
flowchart TD
    SPA["app.example.com — Cloudflare Pages (React SPA)"] -->|"/api/* — same-origin Worker route, no CORS"| W

    subgraph W["contentforge-api — Cloudflare Worker (Hono, 40+ routes)"]
        AU["Auth — D1 sessions · salted SHA-256"]
        CO["Content — articles / battlecards / sources + RSS cron ingest"]
        ME["Media — R2 proxy · presigned + worker-proxied uploads"]
        GE["Generation — Workers AI image/chat via AI Gateway"]
        WF["Workflows — staged pipeline + D1 audit ledger"]
        SC["Scheduler — Postiz client · queue-driven publish"]
        SA["Sales — prospect / outreach / approval queue"]
        RE["Research — SEO + competitor intel (Vectorize)"]
        DO["ScheduleRoom DO — per-user SSE fan-out"]
    end

    D1[("D1")]
    R2[("R2")]
    KV[("KV")]
    QU[("Queues + DLQ")]
    VE[("Vectorize")]
    AIG["AI Gateway → Workers AI · Gemini · Replicate"]
    PZ["Postiz — self-hosted publisher"]

    W --> D1 & R2 & KV & QU & VE
    GE --> AIG
    SC --> PZ

Loading

The Worker treats Postiz as a black-box publisher behind an API key, so it survives Postiz outages and can be swapped for another scheduler by changing a single consumer file. Everything else — auth, media, drafts, schedules, articles, workflows, audit ledger, brand profiles — lives in Cloudflare-native primitives.

Repo layout:

web/       Vite + React SPA  (src/lib/api.ts is the single typed API client)
worker/    Cloudflare Worker (src/index.ts mounts every route; src/nodes/* are pipeline stages)
infra/     D1 migrations (0001–0007) + admin/content seeds
tools/     Python image→video generation script
docs/      Deploy runbook, architecture notes, rebrand guide

🧑‍⚖️ The Sales Approval Gate

The sales workflow is compliance-gated by construction — public sources only, and nothing sends without a human:

%%{init: {"theme": "base", "themeVariables": {"primaryColor": "#0D1622", "primaryTextColor": "#EEF3F7", "primaryBorderColor": "#3B9EF3", "lineColor": "#3078B4"}}}%%
flowchart LR
    D["Prospect discovery — public sources only"] --> EN["Enrichment"] --> OD["Outreach draft"] --> HA["🧑 Human approval"] --> SQ["Send queue"]

    classDef gate fill:#0D1622,stroke:#8EC7FF,stroke-width:3px,color:#8EC7FF
    class HA gate

Loading

🧠 Design Decisions

Decision	Why
Same-origin /api/* Worker route	No CORS, no preflight tax, one origin to secure.
Postiz as a black box	Per-platform OAuth is someone else's full-time job; a single consumer file keeps the scheduler swappable and outage-tolerant.
Audit ledger with input/output hashes	Any pipeline run can be reproduced and inspected step by step — provenance for generated content.
Safety review before surfacing	LLaVA captions + Llama Guard classification gate every generated asset; low scores can regenerate via delta prompts instead of silently shipping.
Durable Object per user for SSE	Live generation status fans out per user without polling.
KV for hot caches	Brand profile and 24h SEO research caches keep repeat reads off the LLM path.
Human approval on outbound sales	Drafting is automated; sending is a human decision, always.
Secrets via wrangler secret put only	Keys never enter the repo; wrangler.toml carries placeholders, not credentials.

⬆ back to top

🚀 Getting Started

Prerequisites: Node 20+, a Cloudflare account, and Wrangler (npm i -g wrangler).

# 1. Worker (API) — local D1 + wrangler dev
cd worker
npm install
cp .dev.vars.example .dev.vars        # fill in local dev values (no real keys needed to boot)
wrangler d1 migrations apply contentforge-prod --local
wrangler dev                          # http://localhost:8787

# 2. Web (SPA) — in another shell
cd web
npm install
npm run dev                           # http://localhost:5173, proxies /api/* to :8787

Important

Provider API keys are set as Worker secrets via wrangler secret put — never committed.

☁️ Pre-deploy checklist — Cloudflare resources

Before deploying you'll need to create the Cloudflare resources and fill the placeholder IDs in worker/wrangler.toml:

• Create the D1 database
• Create the R2 bucket
• Create the Queues (+ DLQ)
• Create the KV namespace
• Create the Vectorize index
• Create the AI Gateway
• Fill the REPLACE_WITH_* IDs in worker/wrangler.toml
• wrangler secret put each provider API key

The full production runbook is in docs/DEPLOY.md.

⚙️ Configuration

• worker/wrangler.toml — all bindings, the cron, the route, and non-secret vars. The database_id, KV id, and R2_ACCOUNT_ID fields contain REPLACE_WITH_* placeholders — fill them with your own resource IDs.
• worker/.dev.vars.example and tools/.env.example — copy to the un-suffixed filenames and provide your own keys for local use.

🎯 Scope & Status

Stated plainly:

• Core generate → review → export → schedule loop — runs end-to-end
• Audit ledger, safety review, media library, brand profiles, research, sales gate, planning, chat, health board — implemented
• Video editor and some Intel panels — prototype-grade, labeled as such in the UI
• Multi-tenancy — not implemented; this is a working single-tenant build
• Public hosting — not currently offered; run it in your own Cloudflare account

❓ FAQ

Can I point this at my own brand?

Yes — that's the intended path. Swap the brand profile, palette, and seed data; the walkthrough is in docs/REPLICATE-FOR-ANOTHER-COMPANY.md.

Does it post directly to Instagram / X / LinkedIn?

No — deliberately. Publishing hands off to a self-hosted Postiz instance via its public API, which owns the per-platform OAuth. ContentForge treats it as a swappable black box.

Can generated assets go out without review?

No. Every generated asset is captioned and safety-classified (LLaVA + Llama Guard) before it's surfaced, and low-scoring assets can trigger a delta-prompt regeneration instead.

Which AI providers does it use?

Everything routes through Cloudflare AI Gateway: Workers AI models (gpt-image, chat, LLaVA, Llama Guard, BGE embeddings) by default, with Gemini and Replicate as optional providers.

Is it multi-tenant?

Not yet — it's a working single-tenant build. The honest-labeling rule applies here too: what's prototype-grade says so.

What happens to outbound sales messages?

They stop at a human. Discovery and enrichment use public sources only, drafts queue for approval, and nothing enters the send queue without an explicit human decision.

📖 Glossary

Term	Meaning
D1 / R2 / KV	Cloudflare's serverless SQLite database, object storage, and key-value cache.
Queues + DLQ	Cloudflare's message queues, with a dead-letter queue for failed deliveries.
Durable Object (DO)	A single-instance stateful Worker — here, one ScheduleRoom per user fanning out SSE.
Vectorize	Cloudflare's vector index — powers brand-profile similarity and competitor battlecards.
SSE	Server-sent events — the live status stream into the generation grid.
AI Gateway	Cloudflare's routing/observability layer in front of AI providers.
Postiz	A self-hostable social scheduler; ContentForge's delegated publisher.
Delta prompt	A corrective prompt derived from a failed review, fed back into regeneration.

🤝 Contributing & Security

Issues and PRs are welcome — open an issue first for anything touching the pipeline stages or the publish path. Security concerns: report privately rather than via a public issue. Secrets never belong in wrangler.toml or the repo.

📄 License

MIT — see LICENSE.

---

_{If the pipeline-with-an-audit-trail approach is useful to you, a ⭐ helps others find it.}

_{Built by Jalen Ward · Ward Tech Systems · github.com/builtbyai}