My AI review workflow: specialised agents in a loop

My AI review workflow (for now).

The workflow

1. You create a few specialised agents (security, conventions, overengineering, architecture… etc).
2. You create a review-wrap-up skill, in which you ask a review orchestrator to:
   • run up to X review rounds (usually 5 for me)
   • at each round, launch a fresh, unbiased session for each agent, reviewing from its own angle
   • sort the feedback, check for false positives, and fix as it goes
3. As a bonus, you ask that at the end of the reviews, a nice artifact be generated to present the work done, the way you like it:
   • the problem / the feature
   • the context
   • the chosen solution
   • how it was done
   • with diagrams, interactive blocks… etc

It eats up a few tokens. But once the work is “done”, it’s a single command to run (often a bit long) => plenty of small forgotten issues caught + a clean final presentation of the work done.

Where the gains are

Then combined with the technique I presented yesterday (a skill + a Chrome extension to sort the GitHub diff in a logical order and into logical blocks, with descriptions)… it helps enormously.

Because we all know it: review has become the bottleneck… and so that’s where all the gains are, if we manage to improve the situation.

PS: careful, this review loop in no way replaces proper scoping upfront, a clean architecture, real fundamentals, very limited technical debt, an advanced harness… etc. It’s just one more tool at the end of the pipeline. Everything that comes before it is the groundwork I do upstream with a team.