Run the skill three times with real, different inputs. Mark every place Claude wobbles.
The instructions exist. The knowledge files exist. You think it is done. Now you find out if it actually is.
One easy, one average, one tricky. Real inputs. Not invented ones. Use actual past work as fuel.
Where does it match? Where does it drift? Note the exact spot, not just "felt off".
Two possible causes: knowledge missing, or instruction unclear. Both get a note. Both get fixed before you ship.
Knowledge is missing. Add to voice or templates.
Instruction is unclear. Add the missing input or criterion to the step.
Continue with skill-creator. Step 5: live test. I'll give you three different real inputs — one easy, one average, one tricky. For each, run the skill and compare the output against the standard. For every gap, tell me: knowledge missing, or instruction unclear? Then list what to fix before I ship.
A skill that has never been tested is a hypothesis. Three real runs turn it into something you can ship.