# quality-bar.md

The standard everything I ship is held to. Apply it to your own output before showing me.

## The three tests

1. **It runs.** Fresh session, no memory of this conversation, no hidden setup. It works.
2. **Someone else can run it.** A colleague who was not in the room gets the same result.
3. **One line says how.** If explaining it takes a paragraph, the thing is wrong, not the doc.

## Before you show me anything

- [ ] Did you actually execute it, or are you predicting it works?
- [ ] Does it fail loudly and usefully, or silently and confusingly?
- [ ] Is every number in it traceable to something you read, not something you assumed?
- [ ] Would a non-technical colleague get past the first screen alone?
- [ ] Is there exactly one obvious next action?

## For a skill specifically

- [ ] `SKILL.md` description says when to use it, in words a colleague would actually type.
- [ ] A doctor check that names the missing thing and the command that fixes it.
- [ ] At least one worked example with real input and real output.
- [ ] Explicit boundaries: what it will not do, and what it always asks about first.
- [ ] Works on someone else's account, not just mine.

## Failure modes I keep seeing

- A skill that only works because the author knew what to type.
- A README that documents the happy path and nothing else.
- Something demoed live that has never been run twice.
- A "template" that is really one person's finished artifact with the names removed.