Behind the Build

This is what it actually looked like

One person and six agents built this site. Here's the reality — including the failures, the corrections, and the branches that got burned.

The Reality

The build in numbers

6AI AgentsEach with defined roles and boundaries
1CLAUDE.mdSingle source of truth for everything
8Thought PagesWritten by Content Strategist, built by FED Developer
0Alignment MeetingsVision flows through context, not calendars
Things That Went Wrong

The branches I burned

Nate Baldwin calls it Branch + Burn. I call it Monday. Every failure below taught me something the success couldn't have. The knowledge survived. The code didn't.

What I tried

Static HTML site with inline CSS

C — Create
What failed

Couldn't scale. Adding a new page meant copy-pasting 400 lines of HTML. Design tokens were CSS variables but nothing consumed them programmatically. The site described a design system it didn't actually use.

What I learned

The site must practise what it preaches. If I'm selling agentic design systems, the site needs to BE one. Migrated to Next.js with Tailwind and real token consumption.

What I tried

Letting the FED Developer write content

<N> — Narrate
What failed

The developer agent optimised for structure, not voice. The content was technically correct but read like documentation — no fire, no provocation, no Lincoln in it.

What I learned

Created a dedicated Content Strategist agent with Lincoln's voice deeply encoded. The developer builds what the strategist writes. Separation of concerns isn't just for code.

What I tried

Designing in Figma first, then implementing

I — Ideate
What failed

Classic waterfall trap. By the time the Figma design was 'approved', the code had already evolved past it. The design became a constraint instead of an enabler.

What I learned

NorthStar Prototyping means code IS the design tool. Prototype in code, iterate in code, ship the prototype. Figma is for tokens and components, not page layouts.

What I tried

One massive prompt to build an entire page

I — Ideate
What failed

Context window overload. The agent lost track of the brief halfway through and started inventing content. The result was impressive-looking but strategically incoherent.

What I learned

Nate Baldwin's 4-step structured prompting was right: context priming first, then planning, then refinement, then implementation. Never skip straight to build.

The quality of input is equal to the quality of output. But the quality of correction is what makes it yours.
The human in the loop
Human Correction

What the agent wrote vs what I shipped

The agents handle volume. I handle voice. Here's what that looks like in practice — raw agent output on the left, what actually shipped on the right.

Competing Values page introduction
Agent output

Quinn and Cameron's Competing Values Framework provides a useful lens for understanding organisational culture dynamics and their relationship to design system implementation methodologies.

What shipped

Your design system team has a culture problem. Quinn and Cameron's Competing Values Framework explains why your team is grinding — and the INC framework maps the way out.

Why I changed it

The agent wrote like an academic paper. I needed a diagnosis that makes a design system lead think 'I've been thinking about my team wrong.'

Agent team page — the economic comparison
Agent output

AI agents can significantly reduce the cost of building digital products by automating tasks traditionally performed by multiple team members.

What shipped

This is not about replacing people. It is about what becomes possible when one person has the right methodology and the right tools.

Why I changed it

The agent went for the obvious 'AI saves money' angle. That's threatening, not inspiring. The real message is amplification, not replacement.

The Recursive Proof

Building the site that describes the methodology, using the methodology

The CLAUDE.md file that governs agent behaviour was itself written through agent conversations. The thoughts pages that describe I<N>C were built using the I<N>C process. This page about failures was reviewed and corrected by the same human-in-the-loop process it describes.

This is not a bug. It is the point. The methodology proves itself by being used to describe itself. If the site reads well, the system works. If the agents stay on-brand, the governance works. If the content provokes, the voice encoding works.

You are reading the proof right now.

Want to see the raw conversations?

I'll walk you through a real build session — the prompts, the agent hand-offs, the corrections, the burns. No polish, no editing. The real thing.

/