Kojo Osei

 undo All posts
Subscribe

How to transfer craft to agents

• 2 min read

A style reference in Midjourney captures the feel of an image. It doesn’t copy any content. It captures things like texture, lighting, and color palette.

Suppose I love the image on the left and want to create something that feels similar. I copy its style reference, write my prompt, and generate a new image with the same essence (see right).

(Left, original): a woman reading a book | (Right, +sref): walking in the garden at sunset

In this case, I don’t care about the original subject. I care about the tacit qualities: the mood and visual language.

Style references are obviously useful in image generation. But I think the idea extends much further.

A lot of knowledge work relies on tacit knowledge. It’s the craft embedded in the output. You recognize good work when you see it. But if someone asks you to explain why it works, you hesitate. There’s no checklist. No explicit playbook. 

So how do you transfer that tacit knowledge embedded in your team to your agents?

Agent skills feel like an early answer. They act as style references for work. 

Take Claude Code. The base Opus/Sonnet models are great for general coding tasks. But you can extend it with skills that embed procedural and stylistic patterns that are unique to how you work.

I’ve found agent skills are strongest when extracted from real examples. Starting with finished work reveals patterns you’d miss if you tried to invent rules upfront. My Claude skill for commit messages, for instance, is derived from Linus Torvalds’ Linux commits. I built a pipeline that analyzes a repo and surfaces the implicit structure in a contributor's commit message style.

This pattern of not writing skills but mining them extends beyond code. Variant UI does something similar for design. It transfers the style of one interface onto another by analyzing the finished artifact rather than relying on handcrafted prompts.

The power in these examples comes from the fact that you don't set out to write a series of long, detailed prompts but that you derive the prompts from finished work. Thus, you capture a bunch of tacit knowledge you may not have noticed or intuited.

I suspect style references, implemented as reusable skills, will become a primary way we work with agents. They encode institutional memory. They preserve craft. And they do something that prompts alone rarely can: they carry the feel of good work forward, even when no one can fully explain why it’s good.