RFCs in the Era of LLMs: Takeaways for Engineers
In RFCs in the Era of LLMs I argued that agents flipped the RFC’s job: writing the document is no longer how you think, and collaboratively reviewing the document is how your team keeps understanding the system it owns. This post is the operator’s half - what changes when you’re the one producing the document with an agent. Takeaways for Leaders covers the other half.
If you take one rule, take this one: never make a human read what you didn’t. The document cost you minutes; you’re about to ask five people for an hour each. Reviewer attention is the scarcest resource in the entire process, and AI slop isn’t a style problem - it’s text nobody chose. Everything below is that one rule, operationalized.
Write the bullets with the laptop closed. The model answers questions; it doesn’t get to decide. The honest accounting is uncomfortable: after hours of back and forth, most of the words in the doc are the model’s, and your irreducible contribution is the decision - what we’re doing, why, what we rejected. So this is the test I run before circulating anything: close the laptop and write those bullets from memory. If you can’t, you didn’t haven’t made any decision yet.
Route the argument through the file. The session dies; the file survives - so the file is where the argument has to live. On my last RFC, reviewers’ objections went into a comment file, and every reply went back into the same file - mine labeled as mine, the agent’s as the agent’s. The next review round, and the next session, started from that artifact instead of anyone’s memory. The unsolved part is the dance: folding the answers back into the RFC itself without losing the session you need to act on them. Workflows like superpowers close some of the gap; I haven’t seen anyone close it fully.
Mark provenance in the margins. My comments on my own last RFC include “the agent’s proposal, not mine” and “I did not say that,” on a paragraph where the doc had quietly attributed the model’s position to me. Reviewers argue differently with a human conviction than with a generated one. They’re entitled to know which is which.
Restore the confidence gradient. Human writing carries confidence levels - “definitely X,” “I think Y,” “no idea about Z, help me” - and reviewers use them to steer their attention. Generated text is smooth: every sentence sounds equally sure, so the reviewer has no map of where to attack. I tried fixing this with formal tags on every claim - verified-in-code, spiked, proposed, open question. Re-reading those docs now, the tagged text is its own kind of slop: technically precise, not human readable. What works is boring, plain sentences: “I haven’t verified this.” “I don’t understand this part, or how it ended up in my RFC.” I left that second comment on my last RFC. It could have been embarrassing; instead, it turned out to be the interesting rabbit hole of the whole design, which got a proper explanation only because I admitted I was lost. When an agent co-writes your document, your own confusion is the best review signal you have. Give your reviewers the map made out of your gut feeling :)