Jev Daily

Moritz Kremb gives every agent skill a feedback doc

One file, read on every run, holding the corrections the skill would otherwise forget.

Moritz Kremb says he now builds most of his agent skills with a built-in feedback loop, and the loop is one document that is read every time the skill runs. When he has feedback about the output of a skill, it goes into that doc rather than into the skill definition. He calls it the single biggest thing you can do to improve your agent skills.

Moritz Kremb
@moritzkremb
X
basically one doc that is read every time the skill runs
Nov 7, 2026 · View on X

The mechanism is deliberately boring. The skill file stays as written, and the corrections accumulate next to it in a place the skill is guaranteed to read before it does any work. Kremb's framing is that the skills self-improve over time, which is accurate in the narrow sense that the behaviour changes without anyone editing the skill, and generous in the sense that a human is still writing every line of the improvement.

What is actually being claimed

This is one builder's working practice, posted as a practice. There is no eval, no before and after number, no count of skills he has done this to, and no example of the document itself. Nothing in the post says how much the outputs improved or on what task. Treat it as a pattern worth trying for an afternoon, not as a measured result.

For people building on it, the cost lands in two places the post does not touch. The doc is read on every run, so every correction you add is tokens on every subsequent invocation, and a feedback file that grows for a few months is a context bill and a latency tax before it is anything else. Second, append-only notes drift into contradiction. Two corrections written weeks apart against two different outputs can disagree, and the model gets to pick. If you copy this, the thing to build alongside it is a reason to delete lines, whether that is a periodic rewrite of the doc or a rule that a correction replaces rather than joins the ones before it.

The pattern is cheap enough that the honest test is your own skills. Pick the one you correct most often, give it a doc, and see whether you stop retyping the same note.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.