An automated creative production pipeline

A good brand system takes a designer days: research the aesthetic, build a moodboard, choose a palette and type, draw icons, write usage guidance, package the files. I wanted to encode that process, keep human taste at the decision points, and let automation do the production. AI spend could only happen on ideas I had already approved.

Each kit was a row in Airtable moving through stages, and every stage was an n8n workflow. Four checkboxes were the entire interface:

• Approve A. Claude researches the aesthetic and writes the creative direction; after my approval, Replicate generates about nine assets.
• Approve B. The pipeline writes the brand primer and packages the full kit, including the coder bundle.
• Approve for store. A preflight check validates every asset and shows me the listing before it exists.
• Push live. It creates the Stripe product and publishes the listing.

• A human approval before every paid stage.
• Claim-first: each paid workflow marks its record as taken before any API call, so a crash parks the record instead of re-billing it every minute.
• A circuit breaker that halts everything and alerts me on any error.
• Validation gates that refuse to make a paid call when an input is missing.

• Chronic halts. Failure alerts hit 900 in about a week. I traced it to the server running ten jobs at once on a single CPU, starving the code runner. Capping it at three fixed it.
• Silent failures. A step reported success while doing nothing because one node was disabled. I found it, and now treat "succeeded instantly" as a red flag.
• Retrofitting the catalog. When the product shifted to developers, I regenerated the coder bundle for every existing kit without touching the originals.

92 brand systems researched, designed, packaged, and published from a spreadsheet, with a person approving every step that cost money.