AI agents (Claude, at least) seem to have this funny inverted writing style, where everything is like:
Conclusion: right at the top
Here’s the middle part, because of the beginning part.
A list (the reason for the list):
- Second thing
- First thing (do this first)
I’ve started putting the following in all my CLAUDE.mds, and, in my experience, it improves Claude’s writing substantially:
Explanation Style
When explaining something, please put premises before conclusions and causes before effects whenever possible. Please define terms and acronyms that I’m unlikely to know before using them and briefly describe systems that I’m unlikely to be familiar with before referencing them. If you’re unsure whether I’m familiar with a particular topic, please ask, and record a memory to inform future judgment about what to explain (noting that my knowledge will change over time).
Please try to give the highest-level explanation that answers the question, avoid repetition, and rewrite explanations if starting from a different place would reduce forward references and other inversions, which are confusing, or would otherwise make the explanation simpler or shorter.
(I want to incorporate this into a longer series about how a tremendous amount of tacit knowledge on effective communication and upwards management is now recorded in CLAUDE.md files around the world. But I felt this tip was useful enough by itself to be worth sharing directly.)