~/projects/agency-delivery-system $ cat README.md

Making an agency's fulfilment agent-ready

I'm helping a small B2B agency in Germany take its client fulfilment, from the first workshop to a live campaign, which already works well, and make it agent-ready.

context
Part-time job, B2B agency, remote
when
2026
status
Ongoing
  • Claude Code
  • Python
  • LLM APIs
  • Google Workspace APIs
  • CRM & email tools

// the problem

Every new client goes through the same steps: a kickoff workshop, a summary for the client, campaign drafts, a review round, then go-live. A lot of that was already automated with AI, but it grew fast and by hand. Nobody could quite say which step used which prompt, model or key, or what a run actually cost. And when something broke, it was hard to tell where.

// what I'm building

First a map: every step with its trigger, input, model, prompt, keys, output and cost, all on one page. Then I rebuilt the whole pipeline in my own lab, with my own keys and test data, so I can take it apart without touching anything live. Now I'm making it better one step at a time: clearer prompts, less overkill, and a check on what each step puts out. The goal is a clean interface for the humans using it, but mostly a system that's ready for AI agents: every step with a clear input and output, so an agent can run it and a person can follow along.

// how I think about it

Pretty much my manifesto in practice: specify first, keep it lean, and give every part a standard slot so a model or tool can be swapped without the whole thing falling over. A car, not a soapbox.

// where it's at

Honestly, still in progress. But the map alone made the whole thing a lot easier to talk about, and the lab means I can try stuff without breaking anything for a real client. It also teaches me a lot about security and handling client data properly. No client data on this page, for obvious reasons.