Local AI: Architecting & Implementing Workflows
Boris Cherney, Anthropic’s Head of Claude Code, explained the agentic paradigm:
“I don’t prompt Claude. I have loops running that prompt Claude and figure out what to do. My job is to write loops, and this is the transition we’re going to see for the rest of the year.”
This article, and much of the Content To Cash series, is about how to go from prompting to engineering loops. ‘Loop’ is just another term for workflow.
A single prompt can hide significant complexity, and we must deconstruct each prompt to understand what skills and context are required to serve the intent. This is how we start engineering loops. I do this exercise in my executive-level seminars to explain why so many agentic initiatives fail and how to fix it.
I ask the audience for a prompt they or their teams have tried to use that doesn’t deliver the output they need. For this article, let’s start with one of Cici’s (Content To Cash agent) core prompts:
“Create an effective LinkedIn post in my voice to market my AI Product Management Certification course.”
Even Claude 4.8 or GPT 5.5 can’t support this without additional skills and context. The prompt is a surface-level description of a very complex workflow. A domain expert in marketing with extensive social media experience would be able to serve my intent by executing the loop.
They possess the skills and context that an LLM alone is missing. BUT, LLMs don’t just generate content. They can use their understanding of language to interpret complex instructions and complete the loop.
In the agentic framework, LLMs can use a skill just like they use any other tool. They can use context about how to use the skill just like they can accept context about how to use any other tool. In this article, I will explain how to determine which skill-context pairs are necessary and how to implement them in an agentic workflow.
Turning Prompts Into Workflows Or Loops
The prompt reveals gaps in the LLM’s understanding, and this breakdown is how I begin the agentic architecture process. We must extract context at the workflow and skills levels so that agents can be successful.