I asked codex for pros and cons of 2 relevant repos i have vs karpathy gist. Will certainly take some ideas!
---
Net: Karpathy’s gist is a strong idea for the memory/synthesis layer Unclawg is currently missing. Unclawg is stronger at review and execution governance. The best version of Unclawg probably combines both: API-backed governed workflows plus an LLM-maintained, continuously compiled knowledge layer over every track.
___
followed up by a variation of
@doodlestein 's "accretive" prompt: "perform /build-vs-clone analysis wrt these gaps, then determine what the single smartest and most radically innovative and accretive and useful and compelling addition you could make to the project at this point to get us on the right track?"
response:
The single highest-leverage addition to any AI agent governance system: Governance Memory.
A compiled knowledge layer that treats every human review decision as a training signal and compounds it into institutional intelligence.
Not a generic wiki — a governance-derived one that only works if you own the decision point.
Why: you already have the most valuable data in the agent stack — what humans actually approve, reject, request changes on, and why. No orchestration framework or execution runtime has this signal. Today it's consumed once during review and archived. The move is to point compounding inward at governance itself.
What it looks like:
Per-track: after each review, LLM-maintained briefings — decision rationale, open questions, risk register, quality drift. Reviewers see compiled intelligence, not raw revisions.
Cross-track: decision precedents, agent quality profiles, reviewer pattern extraction, evolving standards derived from real accept/reject behavior.
Feedback loop: compiled standards feed back into intake. Agents learn "based on 47 past reviews, this operator expects X, Y, Z" before submitting — quality improves pre-review.
Why it compounds: every review makes the next one faster. Every agent submission gets smarter guidance. Institutional knowledge survives any individual agent, runtime, or reviewer.
The moat: anyone can build a review UI. Nobody else has the governance decision corpus to compile from.
----
"deeply consider the landscape between skillbox and unclawg and where this should sit"... hmmmmm
___
Will be lots of fun variations of this popping up over the next few weeks!
Wow, this tweet went very viral!
I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs.
So here's the idea in a gist format:
gist.github.com/karpathy/442…
You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.