Pierce Brand Agency.

Context

Most personal-brand sites in 2026 look the same. Black background. Electric blue accent. Inter font everywhere. "I help X with Y" headline. Linear-clone aesthetic.

That genericness is mostly the LLM's fault. Asking "write a hero section" produces the same hero section everyone gets. The voice averages out. The design averages out. Every AI-founder site converges on the same template.

The fix isn't "stop using AI." It's "use AI with structure that prevents averaging." That meant building an agency, not a chatbot.

Approach

Project-scoped Claude Code subagent system at .claude/. Five specialists, four skill domains, five slash commands, hooks that capture learnings between sessions. Modeled on the 15-agent pierce-agency system I run inside PREP, but slimmed for personal-brand work.

Outcome

This site. Estate Modern visual system: obsidian, aged bronze, celadon sage, bone. Modern serif display, Inter body. Voice rules enforced at every section. Every PR runs through the Reviewer agent before merge.

The argument the brand makes: real estate (warm, established, gold) meets AI (cool, emerging, sage). The colors literally embody the past, present, and future arc of the page.

Why this matters for clients: the same agent pattern can be reused for any brokerage that wants to ship technology without losing their voice. It's not a one-off site build. It's a reusable framework.

Reflection

The hardest part wasn't writing prompts. It was figuring out which decisions to bake into the system versus leave open for the human to pick. Bake too much and the agency becomes rigid. Leave too much and you're back to averaging.

The middle path: bake voice rules and design tokens. Leave content judgment and orchestration to the human. The agents do the work. You decide what work matters.

A few specific lessons:

• One quality gate beats three. PREP's 3-stage QA → Security → Code Review chain works for a complex SaaS. For a personal-brand site, splitting it that way is overhead. Single Reviewer covers everything.
• Voice rules are gospel. The Storyteller has the banned-phrase list embedded in its system prompt. The /voice-check command runs the same grep. Belt and suspenders.
• Knowledge base + incident log per skill. Knowledge base is what worked. Incident log is what didn't. Both grow over time. Surprises pruned, repeated mistakes caught.
• Hooks are where the learning lives. SubagentStop auto-captures learnings before the context vanishes. Without that, every session starts from zero.

Compare this to SNKRGarage, the pre-AI venture. Same engineering instinct, different era. SNKRGarage was figured out by hand, no LLM looking over my shoulder. Pierce Brand Agency is what happens when you bring AI into the workflow but keep the engineering rigor.

Stack

Claude Code · Anthropic Claude Sonnet 4.6 · Vercel AI SDK · Custom Claude subagent framework · Project-scoped agent definitions · JSON hooks (SessionStart + SubagentStop) · Markdown-based skill knowledge bases. Built on the same Estate Modern design system the agency itself enforces.