A brand-locked, compliance-gated content engine for real estate — built as a Claude Code skill.
Point it at one agent's brand folder, drop listing photos into a folder, ask for a job in plain English — and get finished, channel-sized social assets: stills, a narrated video walkthrough, a typography-led sales brief, a teaser cut, and an interactive scroll-through listing microsite. Every asset carries the agent's real logo, exact brand colors, license number, and required disclosures — placed by deterministic scripts and verified by a lint that can actually fail, never left to a generative model's memory.
The Interactive Scroll Tour: scroll-scrubbed camera flights through the listing's real
photos, with doorway-portal transitions. Try the live demo →
Real estate agents fail at consistent content for a boring reason: the production pipeline — gathering assets, remembering brand standards, resizing per platform, compliance lines, captions, scheduling. Generic AI tools produce off-brand slop because they have no persistent brand context and no quality gate. Content Foundry fixes both with three design rules:
- Composite, don't request. The logo file's actual bytes, the exact brand hex, the license
number, and the disclosure line are layered on by
composite.pyat channel-spec coordinates. A generative model is never asked to reproduce them. - The lint must be able to fail.
brand_lint.pyre-reads brand truth from disk and verifies the pixels — wrong hex fails, missing disclosure fails, sub-AA contrast fails. CI proves the failure paths stay wired. - Foundation once, content forever. Each agent is one portable folder (voice, colors, logo, compliance). A brokerage with 40 agents has 40 folders — tenancy is the filesystem.
| Product | Engine | Example |
|---|---|---|
| Channel stills (IG/FB/LinkedIn/Stories…) | Pillow composite + brand lint | docs/media/gallery-*.jpg |
| Digital Walkthrough — narrated room-by-room video | Remotion | docs/media/walkthrough.gif |
| Digital Sales Brief — typography-led 25s video | Remotion | — |
| Listing Teaser — 10s hook cut | Remotion | — |
| Interactive Scroll Tour — scroll-scrubbed listing microsite | scroll-world engine | live demo |
| Draft-first scheduling | Post Bridge CLI (optional) | — |
Requires Claude Code. Then:
git clone https://github.com/BuildThingsThatBuildthings/content-foundry
cd content-foundry
python skill/content-foundry/scripts/doctor.py # tells you exactly what to install
pip install -r skill/content-foundry/requirements.txt
cp -r skill/content-foundry ~/.claude/skills/content-foundryOpen Claude Code in the repo and try the bundled fictional demo:
Run a new listing announcement for the agent in
demo/agents/riley-shoreusing the photos indemo/drop-lakehouse.
It will inventory the drop, write a reviewable creative brief, composite branded stills for each
channel, lint them, and export. For video: cd skill/content-foundry/engines/remotion && npm install once, then ask for a "digital walkthrough."
Setting up a real agent: say set up my brand in Content Foundry — a guided interview builds
the agent folder (see skill/content-foundry/INSTALL.md for the non-technical version).
drop folder ─→ 1 INGEST ─→ 2 UNIFY ─→ 3 RESEARCH ─→ 4 BRIEF ✋(review) ─→ 5 ROUTE
│
┌─ stills (Pillow) ── 7 COMPOSITE ⚙ ── 8 LINT+QC ── 9 EXPORT ─┘
├─ Remotion video (Walkthrough / SalesBrief / Teaser)
└─ Scroll Tour microsite (scroll-world engine)
└──→ 10 PUBLISH (drafts only, optional)
Full details: docs/architecture.md ·
skill/content-foundry/SKILL.md (the workflow itself)
Deterministic stages are Python scripts (stdlib + Pillow). Creative stages are the model's job. The line between them is the product.
| Tier | Needs | For |
|---|---|---|
| Core (stills) | Python 3.10+, Pillow | Everything in the stills pipeline |
| Video + tour | ffmpeg, Node 18+ (npm install in the engine, once) |
Walkthrough/SalesBrief/Teaser, Scroll Tour |
| Optional | Post Bridge account + API key | Draft-first scheduling |
| BYO | Any TTS + music source (schema in references/engine-remotion.md), any image-gen API |
Narration, music, generated base imagery |
scripts/doctor.py checks all of it with per-OS install hints.
- Image generation is bring-your-own. The composite/lint pipeline consumes any base image file; a built-in generation call is not implemented yet. (The demo photos were generated with an external tool and are labeled fictional.)
- Audio is bring-your-own TTS/music via a documented json contract; no bundled voices.
- Windows is doctor-checked but less battle-tested than macOS/Linux — reports welcome.
- Roster/batch mode (one prompt across N agent folders) is specified, not yet built.
The Interactive Scroll Tour of the fictional demo listing is published on GitHub Pages
(gh-pages branch) — open the live walkthrough →
Scroll to walk room-to-room through doorway-portal transitions; every camera flight is generated
from the demo's still photos with ffmpeg, no AI video.
To rebuild it yourself:
python skill/content-foundry/engines/scroll-tour/build_tour.py --agent demo/agents/riley-shore --scenes <your scenes.json> --out tour/
cd tour && python3 -m http.server 8000 # open http://localhost:8000Content Foundry enforces that required language (license number, brokerage disclosure, Fair Housing line) is present and legible on every asset. The brokerage supplies the wording; the skill never drafts or paraphrases legal text, and never auto-publishes. This is assistive tooling, not legal advice. The prompt library additionally encodes Fair-Housing-aware copy rules (describe the property, never the buyer).
Every property image in this repository depicts a fictional home and was AI-generated for demonstration — because Content Foundry's own rules forbid generating imagery of real listings. Riley Shore and Shore & Co. Realty are fictional.
- scroll-world scrub engine by oso95 (MIT) powers the Interactive Scroll Tour.
- Pretext by Cheng Lou (MIT) computes deterministic text layout in the video engine.
- Remotion renders the video products (note Remotion's own license terms for larger companies).
- Playfair Display & Source Sans 3 fonts, SIL OFL 1.1.