When most people start building AI into their agency, they go straight to the tasks. Pick a process, write the steps, get it running. That's a great way to start, and it's exactly how it should work early on.
But as you scale and start adding more workflows, things tend to break down. The AI loses context it should already have. Instructions start contradicting each other. The same information gets re-explained in every file you create.
Before investing too much time down that path, it's worth stepping back and thinking about the overall structure that needs to be in place for a system like this to hold together long term.
There are three layers of information that need to be organized in order to build AI automations across the different functions of your agency.
Layer 1: Universal (company handbook) This is everything that's true about your company regardless of function or task. Identity, services, pricing, standards, org structure, communication style. It gets written once and shared across the entire system. Every workflow you ever build inherits this context automatically. When something changes at the company level, one update propagates everywhere.
Layer 2: Functional (agent guidelines) This is everything that's true about how a specific function or role operates. The tools used, the formatting standards, the quality bar for all work in that function, how tasks get logged and tracked. It assumes the company handbook is already loaded, so it never re-explains what's covered there.
The rule of thumb: if the same instruction would need to appear in every workflow you build for this function, it belongs here instead. Write it once in the agent guidelines, and every skill underneath inherits it.
Layer 3: Skills (task steps) This is the step-by-step procedure for one specific task. What to fetch, what to do, what to produce, where to record it, how to close it out. It assumes both layers above are already loaded. It contains only what's unique to this one workflow.
This is the layer that changes most often, and that's fine. Because the context above it is stable, updates here are surgical. You're refining one workflow, not rebuilding the whole system.
The classification test
Before putting any piece of information into your system, run it through three questions:
Is this true for the whole company, regardless of function or task? It goes in the universal layer.
Is this true for every task in this function? It goes in the agent guidelines.
Is this only true for one specific task? It goes in the skill file.
The shortcut: if the same instruction keeps getting copied into multiple skill files, it belongs one layer up.
Why this matters
Most agencies that start building AI workflows will figure out the skills layer on their own. That part is intuitive. Fewer will think about the structure underneath. Getting these layers right early means every new workflow you add makes the system stronger instead of more fragile. Skipping it means eventually hitting a wall and having to reorganize everything from scratch.