Skip to main content

Thoughts on AI workflow development

30 June 20262 min readen
Linards Berzins

Linards Berzins

Experienced client-facing HTML developer with a proven track record of delivering highly personalised email campaigns for global digital marketing agencies and top-tier brands.

Thoughts on AI workflow development

Some thoughts on building products with AI. Would be great to read people's ideas on this.

For a couple of years the skill that mattered was prompt engineering.

Trying combinations of words and tech specs until the model did what you intended it to do. Since I started to build agents, I've noticed a trend in agentic workflows, where the model in the middle now matters less than it did before.

There are a few new concepts I'm learning to master in steering LLMs, intent engineering and harness engineering, that seem to have dethroned prompt engineering.

Intent engineering is removing the micromanaging of tasks and the how-to of achieving your goals. Telling the model to do something step by step is mostly gone. The idea is to say what you want clearly enough that the agent works out the how. The goal is to precisely lay out what "done" actually means. It reads fairly simple. However, it is not, definitely not for me. Most of us don't know what we want until something forces us to put it down in writing or go through some sort of detailed questionnaire. To narrow down the idea of a feature or technical spec, I always ask AI to ask me as many questions as possible. It takes longer. However, it's way worth it. The plan is way better when I'm being questioned by agents, in my experience. There is the intent. What do you actually want?

Harness engineering is everything around the model. The tools it can call, the context you feed, the loop that lets it check its own work. Setting up an agent harness takes time too, as you have to choose the right Skills, create subagents if needed, MCP servers, built-in hook guardrails, to name a few.

I've built a couple of dozen skills now for AI applications, and the more I used Matt Pocock's /teach skill, the better I understood both concepts, or any other topic. It's my process now to include the /teach skill in my agentic workflow.

To conclude, the model is swappable. Intent and harness are what I keep.

Share:

Comments