AI systems that give you back the week.

An AI publishing harness for a solo education business

A solo publisher can hold the thesis and still lose the week to everything around it: finding a topic with demand, checking the angle, sourcing claims, generating images, formatting for Substack, and keeping the voice consistent. We built an AI harness that does that work and stops at the decision.

It researches demand and seasonality, drafts in the publisher's voice, checks structure, sourcing, and layout, generates and optimises images, and leaves a formatted draft in Substack. AI handles the open-ended work; scripts handle the repeatable work and the guardrails. Missing positioning stays missing, invented experience fails, and a 22-check guard blocks writes to the wrong publication on a shared account. Publishing stays a human action. The live publication is at forgotteneducation.substack.com.

Read the full case study

A publication that AI agents can actually read

Most sites answer an AI agent with a full HTML page, a cookie banner, and a wall of navigation. We built markdown negotiation into this publication: when a request asks for text/markdown, the site returns the article with its frontmatter, and advertises the alternate in the page head and HTTP headers.

One site now reads cleanly to people and machines: 48 posts converted, Level 3 Agent-Readable in an external scan, with tables, code, captions, and nested lists preserved. HTML stays the default for everyone else, and one plugin covers every post instead of a hand-built sample.

One canon, three products: books, curriculum, shop

A 26-book alphabet series, a homeschool curriculum, and an Etsy shop can drift into three businesses. We built a file-based operating system that keeps them one: locked canon for the series, one brief per project, and a shared voice that survives every new letter, lesson, and listing.

The system generates art with image models, styles it, builds print-ready PDFs, and turns finished units into shop products. The publication is the front door we are building; the books and curriculum are the deeper offer. Same owner, same business, one source of truth.

An AI harness built to spend tokens like money

Most AI systems get expensive because every step pastes its work back into the conversation. We built the opposite. Subagents write to files and return a path, state lives in files, and only judgment calls reach a model. Formatting, resizing, scheduling, and tests run as scripts. If a script can do it, a model never sees it.

The same architecture guards quality. Sources are checked, voice drift is measured, layout is gated, and a human makes the publish decision. It costs less to run and is harder to degrade as it grows. Cheap is not the point. Durable is.

View the reference implementation