Agent skills should be evergreen — continuously improved, not one-and-done
The first version of a skill is never the final version. A dry run might produce an okay first effort — maybe 50% of the way there — but the whole point is that skills should be evergreen. Every time a skill is used, it can improve: either automatically, through the LLM generating learnings and suggestions to improve itself, or interactively, because you're looking at the content and saying 'that's not quite right — here's why you got that wrong.' Over time, this feedback loop continuously raises the performance of the agent.
Go deeper with Greg Isenberg
Ask anything about this, and Greg Isenberg answers from everything they've shared.
More to learn
See everything in Learn
MembersStiff AI-generated content can be fixed in real time through skill feedback loops
When reviewing AI-generated tweet drafts, Greg Isenberg noted a common quality problem: the output felt too formal and stiff rather than sounding friend-to-friend, which is the tone that tends to perform well on social media. The fix isn't
MembersLLM-as-judge: automated quality scoring for agent outputs
Instead of relying on yourself as the sole judge of every agent output, you can use a separate LLM as a judge that fires off and scores outputs along the dimensions you care about. You can then see the trend line of how an agent is scoring
Why agent observability becomes critical when running at scale
Once you start running agents at scale, secondary capabilities like observability become really critical. It's not just about whether you can get one agent to do one thing — it's about how you oversee and run an entire business with many di