Intention is all you need
I find myself using the term “intention” a lot recently. I’ve been asking questions like:
- an AI agent wrote a summary of the changes in this PR, but what was the author’s intention for these changes?
- this LLM-generated PRD says a lot about this feature, but what was the original intention for users?
- this commit message describes what code was changed, but what was the developer’s intention that was fed to the coding agent that wrote this code?
I find myself basically asking people to show me their prompts. While reviewing this LLM output, it would be very helpful to know what your input was that generated all of this.
Especially when a few sentences written by a human can be turned into
- several pages of documentation… and then
- dozens of tickets in your project management tool… and then
- thousands of lines of code (tens of thousands??)
It’s important for any humans-in-the-loop who must review the downstream output of an LLM to understand the intent of the person who initially prompted the LLM.
Adding to Workflows
I have been playing around with exposing intent in my own workflows:
- PR bodies contain a copy of all of the prompts I sent to agents (when feasible)
- when I share an LLM’s analysis of a PR, I include the prompts I sent to the LLM as well as its output
- when I generate a technical design document, I include the prompts I sent to the LLM at the top
- this shows the reference PRDs, tickets and any manual changes not captured in those (like reducing scope for MVP)
How are you capturing and sharing the intention you are feeding to your AI agents?
