00Paige / 2022 — 2026
A consultancy that became a product
Four years of doing marketing production by hand became the spec. I built the product solo.
01Situation
I founded a writing consultancy in 2022 and scaled it to $200K in annual revenue through cold outbound, referrals, and retention — twenty-plus concurrent clients across finance, manufacturing, logistics, and tech. The work was ghostwritten books, grant proposals, and patent applications: twelve books and counting, each one requiring me to hold a client's voice, vocabulary, and prior claims in my head for months at a time.
02The real problem
The constraint wasn't writing speed. It was that every engagement rebuilt the same scaffolding by hand. Each client had a voice, a set of approved claims, a house vocabulary, and a body of prior work — and all of it lived in my memory and a folder of documents. Nothing carried between projects, so onboarding a new client cost days before a word got written, and consistency across a twelve-month book depended on me remembering what we had already said. The bottleneck was not production. It was that the brand context had no home outside my head.
03What I built
04Mechanism
- Client documents, prior work, and approved claims
- A brief naming the artefact and its audience
- House vocabulary and voice constraints
- A living brand library that persists between engagements
- Drafts that inherit voice without re-briefing
- Graded output, so quality is measured rather than felt
05Result
How: Cold outbound, referrals, and retention across twenty concurrent clients in finance, manufacturing, logistics, and tech.
How: Seven purpose-built agents, thirteen background workers, forty-four migrations, and a custom eval harness with graders — written by me.
How: Each one holding a client's voice and prior claims across a months-long engagement, which is the constraint the brand library was built to remove.
06Decision log
Shipped v1 on Google Sheets, Apps Script, and the Claude API.
The question was whether a living brand library changed the work, not whether I could build infrastructure. A spreadsheet answered it in weeks.
It could not scale past a handful of users, and everything in it was thrown away when v2 started. The learning carried; the code did not.
Built v2 alone rather than hiring or raising.
I had the spec from four years of doing the work by hand, and handing that context to someone else would have cost more than writing it.
Roughly a year of solo build time, and a bus factor of one. Every architectural decision is unreviewed.
Wrote a custom eval harness with graders instead of shipping on vibes.
Generated output either holds a client's voice or it doesn't, and I could not tell reliably by reading a handful of samples.
Weeks that produced no user-visible feature, and a second system to maintain alongside the product.
07Caveat
This is a solo-built product with a small deployment history, not a scaled SaaS business. v1 reached daily use by one marketing team; v2 has not been through the load, the support burden, or the edge cases that a real customer base produces. The engineering claims are verifiable in the repository. The product claims are not yet proven by anyone but me.
08What I'd do differently
The eval harness should have come first. I built it third, after v2's agents were already in daily use, which meant the first weeks of output were graded by reading samples and forming an impression — exactly the thing the harness exists to replace. Having the graders from the start would have changed what I built, not just how I measured it.
TODO — confirm: v1's deployment at PIMS was with YSL's marketing team. Verify the engagement was YSL specifically and that naming them publicly is fine, or cut the client name and describe the deployment generically.
TODO — confirm the 2022 founding date and the "$200K annual revenue" framing. The homepage says "built and scaled solo"; if that figure is a peak year rather than a run rate, the label should say so.