Reviewing work with AI agents
- 3pieces
- 1seller
- $1lowest price
How can a structured agent review catch central proof errors that a plain chat review missed, and what did it still overlook?
In one qualitative n=1 case study, I ran Deep Review on version 1 of a publicly available ML preprint before examining its corrected version. The review calibrated against the paper's goals, split the work into proof, theory-to-practice, and empirical-design passes, used local code to test the theorem, and ran separate critique agents against the findings.
Useful if you want a concrete case of structured AI review catching proof errors that a chat review missed, and what it still missed.
From Dipak’s Working Lore · updated Oct 6
How do you structure a coding agent to review a technical paper rigorously instead of in one shallow pass? The four-stage flow and the de-biasing trick behind it.
Deep Review runs from a coding agent interface (Cowork, Codex). Flow: (1) a calibration interview asks what the artifact is, who the audience is, and where to focus.
Useful if you want a coding agent to review a technical paper in depth, including checking theorems with code.
From Dipak’s Working Lore · updated Oct 6
When you need rigorous artifact review, which medium holds up — general AI chat, manual review, or a coding agent? The tradeoffs, and the one gap that decides it.
The three review options each fall short: general AI chat reviews in a single pass, isn't specialized to inspect an artifact over tens of minutes, doesn't verify claims with code, and doesn't check its conclusions with an independent agent.
Useful if you need a rigorous review of a paper or technical document and are choosing between chat, manual review and a coding agent.
From Dipak’s Working Lore · updated Oct 6