Paste a website URL, get a service-ad pipeline

The product problem

Turning a website into recurring Instagram service ads is a chain of decisions: scrape the page, extract a business profile, write ad copy, generate a background, overlay copy, place a logo, compose slots, schedule posting. Power users can wire that on the canvas. Most people will not.

We wanted one action that feels product-simple ("paste your site URL") and expands into a real, editable workflow they still own afterward.

What the user experiences

1. Create a project with an empty workflow canvas
2. Open Workflow Presets → choose Website Service Ads
3. Paste a website URL and apply
4. Land on a wired canvas: scrape source, business profile, ad copy, image, overlay, tip-card compose
5. Logo, brand styles, and business profile update shortly after (background import)
6. Review outputs, then connect an Instagram poster

Same pattern for other presets (app review video, presenter video, clips, blog): different graphs, same apply model. Website Service Ads is the square-graphic path that starts from a live link.

Presets are a graph DSL, not form defaults

Each preset declares nodes and edges in code: inputs (data source, knowledge files), automations (action + prompt + file type), templates (slot layout), and canvas positions. That graph is the source of truth for both preview and apply.

Product benefit: adding a new preset is mostly declarative. Engineering benefit: preview, illustration, and materialization all read the same config instead of drift between UI and runtime.

Apply: one transaction, then async polish

Applying a preset only works on projects with an empty workflow canvas. That constraint is intentional: merging a full graph onto an existing one is a product and engineering mess we chose not to solve on day one.

Inside a single DB transaction we:

• Create library resources (scrape endpoint, stock/knowledge placeholders)
• Create automations with their prompts and process types
• Resolve / attach the content template
• Create chains (edges between nodes, including template slots)
• Save canvas layout so the graph opens where it should

Then, for website presets, we enqueue a background job to refresh site assets. The user can explore the canvas immediately; branding catches up without blocking apply.

The interesting bit: bootstrapping brand from a URL

A service ad without a logo, business profile, and brand tone still “works,” but it doesn’t feel like their site. So after apply we import:

• Business profile: services, benefits, CTA, audience, and tone from page copy.
• Site logo: crawl the page (and a few related paths like About), collect image candidates, score them (header marks, schema.org, common /logo paths), and let a vision model pick the best match against a prompt like “main brand logo, not favicon.”
• Brand styles: pack visible page text + CSS cues, then ask a text model for short reusable notes (colors, fonts, tone), not a dump of raw CSS.

Heuristics first, AI as judge: that keeps cost bounded and avoids asking vision to score dozens of noisy images. If import fails, stock fallbacks keep the pipeline runnable.

How this differs from Agentic Transform

Same platform building blocks (text-to-image, overlay, merge layers); different product shapes. Presets encode a known happy path; the agent explores within a small tool kit.

What we’d tell another team

• If the workflow has more than a few nodes, ship a graph preset, not a wizard of disconnected forms
• Keep apply atomic; enrich branding async
• Refuse messy merges (empty canvas) until you have a clear product model for them
• Pair heuristics + AI for site assets; always have a fallback
• Let users edit the graph after apply: presets are a head start, not a black box