Posts tagged with "agent-skills" · View all
Gmail's own result count reported 201 for a true count of 35. The output flag exits 0 and writes nothing. An unsent draft passes attribution checking because its id is genuine. All three return success. Here is the skill I built around not trusting any of them.
A re-snapshot and a real diff look identical on the page, and nothing in the report tells you which one you are holding. Here is how I made "changed" a server-side fact instead of a writing-discipline problem, and the four numbers from 32 live probes that would each have put a confident wrong figure into a report.
I built a script that checks whether every number in a pricing memo cites its source, then ran it on my own first draft. 36 numeric claims, 36 unsourced, FAIL. Getting to zero took four rounds and found four defects in the skill itself.
I shipped two skills that published green at 100 out of 100, moved the cards to Done, and only found out five weeks later that no customer could see either one. `tier` defaults to `internal`. That gap is why the Skill Creator profession exists — and why its eighth workflow refuses to call anything shipped until it has fetched the artifact and read it back.
I installed the same skill twice, two documented ways, and counted the files. One gave me a single SKILL.md. The other gave me 166. The single-file version looked fine — it just quietly wasn't going to work. That finding is why @di-atomic/propagate v0.1.0 exists.