Matt Pocock writes about the `/retro` skill, an agent plugin that reviews a coding session's record to suggest improvements to the working environment rather than the code itself. It identifies moments where the agent struggled and proposes fixes like navigation pointers, automated checks, or coding standard rules, categorizing each finding by the type of failure it addresses. The skill operates on the principle that mechanical violations should become deterministic checks while only genuine judgement calls become prose in reviewer standards, and it deliberately keeps humans in the loop by only proposing changes until explicitly accepted.
- The skill is explicitly not a memory system; it changes the environment so mistakes cannot recur rather than storing what happened.
- A known weakness is that it can only evaluate one session, so it cannot detect when previously proposed lint rules or hooks have become outdated or overly noisy.
- The author notes a common failure mode where the agent "tends to forget the struggles from the middle of the session and invents generic advice to satisfy the retro categories."
- The skill is explicitly not a memory system; it changes the environment so mistakes cannot recur rather than storing what happened.
- A known weakness is that it can only evaluate one session, so it cannot detect when previously proposed lint rules or hooks have become outdated or overly noisy.
- The author notes a common failure mode where the agent "tends to forget the struggles from the middle of the session and invents generic advice to satisfy the retro categories."
