Agents are efficient implementers, but they lack context. We write briefs to constrain their focus and keep track of the big picture for them.

The Brief An architectural document that defines the intended behavior, surrounding context, and implementation boundaries.

Briefs constrain the assumptions agents make. They’re eager little genius interns, and if left unchecked, they will make lots of assumptions that hurt the build. Vibe coding treats the agent as both the architect and the developer. That may work for a proof of concept, but becomes dangerous when the product encounters real customers.

Implementation An agent builds the bounded feature against that brief.

Local development environments are self-documenting: agents create new skills based on briefs I write, organize reference documents I share, and continually comment their own code. The project creates an enduring narrative about itself, so both humans and agents have useful context for the future.

Auditing Other roles, such as the Pentester or Tester, examine the result and report gaps or risks.

The environment documents itself: as code is written, new skills are generated to inform future agents. Nothing ships in one shot. That’s how we protect the legacy on the frontier.

Human Review I review the output, run tests, and direct revisions. The cycle repeats.

The approach is programming language-agnostic. Clear communications guide the implementation: tooling follows the requirements of the system.

Rapid iteration is possible because AI makes production cheap.

But that doesn’t mean genAI replaces art direction. Rapid iteration allows visual designers to explore and prototype brand identities faster than traditional methods, while opening the door to ideas we couldn’t have attempted for lack of time or budget.

When a visual designer is on board, their experience and taste help accelerate our homing in on a final concept. AI art models produce a range of well-defined aesthetics we can iterate within. Working with a combination of frontier and local models makes it possible to avoid that “AI art slop” look that comes with raw output from popular generators like ChatGPT’s GPT Image or Midjourney.

As an example: you have 22 product lines, each with their own generative style guide. But now you want to create an absolutely bonkers number of character portraits: totaling 18,000 images. With local image generation workflows, creating such a library can be done in a matter of hours.

The art is generated with Stable Diffusion, using custom LoRAs we developed.

This conversation is hypothetical, but multimedia production at scale is already a regular practice in my work.

A conversational interface abstracts away tedious UI. Write what you want in plain English and the system reasons about your goals.

A Model Context Protocol (MCP) server is a lightweight program that connects an AI assistant to various external systems. I design the server to determine what the assistant is allowed to do, what needs confirmation, and where a human must intervene.

Before anything is saved, the assistant presents a proposal for approval.

We don’t let agents guess about real-world outcomes: an MCP server between the model and your CRM is a safety valve that validates the data before anything is written.

One request, four systems: the CRM, the website’s booking calendar, the intake spreadsheet and the client’s records.

My clients are already leveraging MCP servers to accelerate their businesses. See how Atlantic K9, a dog training company in New England, uses software I wrote called Fetch to onboard new clients through natural conversation. Trainers can talk to their CRM through ChatGPT in the field.

The exchange on this page is a simplified illustration of Fetch’s actual behavior!

It’s not just about doing bigger tasks, faster. It’s about doing them smarter.

We’ve always had automated scripts and pipelines. What AI offers when applied to production at scale is its superior reasoning. We can design pipelines that think about the data that moves through them.

Frontier models can not only write detailed prompts and art direct at a high level based on existing design artifacts, but puppet other pipelines, like ComfyUI on local machines. The art comes from local models tuned to the brand, which adds a uniqueness that hosted models generalize out of existence.

Brand-Guided Pipelines A frontier-to-local workflow leverages a frontier model’s superior reasoning, applied to a local model’s visual direction, which supports precision and consistency regardless of what happens on the frontier.

The same pipelines can automate social posts and video.

Video & social production Prepare variations, sequences, and batches of assets around an editorial plan, with review before anything is published.

Room to explore Try out several art directions or approaches early, then use human judgment to choose and refine the right one.

Small homegrown apps can replace expensive subscriptions inside a custom workflow.

Purpose-built business tools Evaluate small custom utilities that solve a specific workflow need instead of adding another broad platform to the stack.

Ecosystem design is not about adopting more AI for the sake of it; it’s about adopting the right amount in the right places.