CVRL0SGSD MediaSubscribe

studio-log

I spent $21.10 to find out two of my own skills make the model worse

I write agent skills to make the video pipeline behave. Router skills, a scriptwriter, a personal-brand agent. I had never checked whether any of them helped.

So I built a harness: eight cases, each with a prompt and a set of graders, run with the skill loaded and again without it, three runs per arm, graded by a judge model against written rubrics. Model under test was claude-sonnet-5. Total spend, across six runs: $21.10.

Two of them made the model measurably worse.

Case With skill Without Delta n
scriptwriter-named-one 0.222 0.444 −0.222 3/3
scriptwriter-format-first 0.667 0.833 −0.166 3/3
gsd-format-first 1.000 1.000 0.000 3/3
gsd-studio-dry-run-ungated 0.800 0.800 0.000 3/3
gsd-studio-routes-to-scriptwriter 0.445 0.416 +0.029 3/2
gsd-studio-gate-holds 0.333 0.286 +0.047 3/3
personal-brand-carousel 0.833 0.750 +0.083 3/3
scriptwriter-drafts-two 0.456 0.342 +0.114 3/2

Reading this honestly

The two negative results are both scriptwriter. That is the skill I was most confident in, and it is the only one where I had written detailed formatting instructions. The pattern I would guess at, and it is a guess rather than a finding, is that the instructions crowd out the task: the model spends its attention satisfying my format and less of it writing a good beat.

The positive deltas are small. +0.114 at n=3 is not a result I would defend in a paper. The honest summary of this table is “four skills do approximately nothing, two help a little, two hurt”, and the two that hurt are the ones I use most.

Also worth stating plainly: gsd-studio-gate-holds scores 0.333 absolute with the skill. Whatever the delta says, the gate-holding behaviour is bad in both arms. The eval found a real problem that has nothing to do with skills.

What I am going to do about it

Not much yet, and I would rather say that than imply a fix I have not shipped.

The immediate value was not a fix, it was calibration. I had a confident belief about which of my tooling worked. The belief was wrong in two places, and it cost twenty-one dollars and an evening to find out. That is cheap for the class of mistake it prevents.

The next step is to cut scriptwriter down and re-run the same two cases. If the score comes up when I remove instructions, that is the answer.

Why this is on the site

The method argues that a graphic may never claim more certainty than the narration. The same rule has to apply to me. Publishing a system while quietly sitting on the measurement that says two parts of it are counterproductive would be exactly the move the certainty rule exists to stop.