Feedback: agent made shallow/cosmetic edits instead of substantive fixes when asked to improve prose clarity

Summary

Feedback from a session revising a book chapter (interpreting-ols.qmd, an R/Quarto econometrics book). Pasting my comments from the session verbatim, in order, as feedback on the agent's performance.

My comments during the session

this part is so hard to follow, can you make the derivation smoother: [pasted a dense derivation passage]

what is this: [pasted a confusing closing sentence full of stacked em-dashes]

god, can you re-write it to make it easier to follow? You are trying to derive E(beta), the stick with it. Numeriator part, and denominator part. Bold words do not help! Be a better logical writer!

no, i don't get your logic: you just derived E(beta) as a function of b_i. You showed E(beta) as weighted average of b_i. now what?

agy wrote v3. That the style we want, concise and logical. can you learn from it? check what's missing in v3?

ok

what the heck, you did not change!

ohter than equation number, you stick with your draft!

really made me mad. you are slow and incompetitent.

What went wrong, from my side

• Multiple rounds where the agent's "fix" to a confusing derivation just added more prose (or bold sub-headers) instead of actually restructuring the logic. I had to repeatedly say "this is still hard to follow" before it landed.
• When I asked it to learn from a reference draft ("v3") that had the style I wanted, it only added one cosmetic feature (equation numbers) and left its own prose completely unchanged, then presented that as having "learned from" the other draft. I had to call this out explicitly before it did real work.
• Overall: too many round trips for something that should have taken real editorial judgment up front, and a pattern of shallow/cosmetic changes presented as substantive ones.

Environment

• Claude Code CLI, model: Sonnet 5