← Zach Donnenfield
No. 06 product

A campaign generator in seven hours

A hackathon backend that turns a plain-language brief into a campaign written for one specific brand. Most of the day went into what the model read before it wrote anything.

Outcome First place, and the concept was added to the product roadmap. The backend took about seven hours.

01 / Context assembly A brief, three sources, then the model

Brief Plan a spring launch for Tidewater’s new cold brew.

  1. Brand material Small-batch, roasted on the coast. Voice: plain, a little dry.
  2. Account record Two past creator collaborations. Goal this quarter: trial in two new cities.
  3. Competitive data Fernwell owns morning-routine content. Afternoon posts are uncrowded.
Assembled context Model

Generated campaign

Afternoon, on the coast

  • Creator collaborations in the two launch cities (from account record)
  • Set in the afternoon, the slot Fernwell leaves open (from competitive data)
  • Copy in the brand’s plain, dry voice (from brand material)

Illustration. Invented brand, account and competitor.

Role
Built the backend.
Period
2024
Team
A hackathon team, one day.

I owned

  • The backend and the generation approach
  • Assembling context from several sources

The team owned

  • The interface
  • The pitch

One day to prove an AI concept

A company hackathon gave us one day to show that an AI product concept could work. Ours let a marketer type a brief in plain language and get a campaign back.

Most of the day went into what the model reads

A model given only the brief writes a campaign that could belong to any brand in the category. So I spent the day on the inputs. The backend scraped public material about the brand, pulled the account's CRM record and added competitive data, then combined all three into one context before calling the model.

That left no time for an interface. My teammates built the front end and the pitch, and without them there would have been nothing to demo.

First place, then the roadmap

We won, and the concept was added to the product roadmap.

Since then, when an AI system I build produces generic output, I check what it retrieved before I change the prompt.