Learning to make AI part of the way I work.

Start with work I already understood.

Pixel & Bracket gave me real problems to work on: uploading content, preparing thumbnails, researching topics, and managing production. My early attempts were graphical tools and dashboards with AI API integrations. Each one helped me learn how to connect an interface, instructions, data, and an output.

As I worked with agents, I started paying more attention to the process around a task. What context does the agent need? What should it produce? Where should that result live? What do I need to check before the next step?

Those questions became as important as the tool itself. I was learning to define the work clearly enough that I could run it again, inspect the result, and improve it.

Give the workflow a structure I can return to.

I moved toward an ICM folder structure with shared instructions, references, stage folders, and saved outputs. Each stage describes its inputs, the work to do, and what it should leave for the next stage. The files hold the context between sessions.

That structure gives me more consistent, repeatable outputs while keeping the process flexible. I can review a result, revise the instructions, or rerun a stage without having to reconstruct the entire workflow in a conversation.

Instructions

Define the task, the relevant context, and the expected result.

Outputs

Save the work where the next stage—and I—can find and inspect it.

Checkpoints

Make review and approval part of the process before work moves forward.

Put the method to work in the business.

My video-production workflow is now central to content creation for Pixel & Bracket. It connects the steps around a recording: preparing source material, reviewing privacy findings, editing, creating a thumbnail, uploading, and reviewing the video before publication.

The workflow carries project files and reports between stages. Agents and scripts prepare work; I make the creative decisions and review the points that need my judgment.

Prepare & review

Bring in the recording and review privacy findings. Redact Pro gives me control over the masks before export.

Edit & package

Work through the video edit and thumbnail with the project’s context and outputs available at each stage.

Upload & publish

Upload to YouTube as unlisted, review the result, then publish and file the finished project.

The checkpoints are part of what makes this useful. A prepared upload still needs review. A suggested redaction still needs a person to check it. I’ve learned to design those decisions into the workflow.

Keep the process portable.

I want to run the same workflow from whichever AI tool fits the work. The instructions, references, and outputs live in the folder structure, so a new agent can pick up the context there. I can change the model or the application running it while keeping the process intact.

That is what model-agnostic means in my work: the workflow has a life beyond a particular chat or interface. A compatible agent still needs access to the files and the tools a stage uses, but the operating instructions travel with the workspace.

This gives me room to keep learning. I can try a different tool, improve one stage, or add a capability while preserving the parts of the workflow that already work for me.