- Answers that read like a person wrote them (0.24.0)
The prompt asked for no emoji, which was a ban, and what was wanted was much
less decoration and answers people are glad to read. The ban also bought
nothing. Measured against the model this product runs, on eight finished tasks
with the answer asked for three ways: the arm carrying no instruction at all
produced ONE emoji in eight.
What it produced instead is the actual complaint. 19 bold labels in front of
paragraphs, 8 numbered checklists, 7 em dashes, 2 openings that announce what
is about to be said, and 2 closings that summarise what was just said. That is
what makes an answer read like a machine wrote it, and none of it was addressed
by a sentence about emoji.
So the prompt asks for the shape instead: the answer first and then what
supports it, no announcing, no repeating the question back, no closing summary,
prose by default with a heading or a list or a table only when the content
really is one, decoration rare rather than forbidden, and every dash a plain
hyphen. Same eight tasks with it: 3 bold, 0 checklists, 0 em dashes, 0
preambles, 0 restatements.
⛔ The gate asserted the WORD markdown where the requirement is a meaning. The
prompt now names the marks it allows - a heading, a list, a table - which says
the same thing to a model and says it more precisely, and the old assertion went
red on the better prompt. It checks the meaning now, and four known-bads kill
it: no sentence on decoration, the marks unnamed, "plain text" back, and an
emoji or an em dash in the prompt itself.
- The sessions control says Sessions
Three lines mean "there is a menu here" only to somebody who has already
learned that they do, and what is behind this one is a list of sessions, which
the column's own header says in words. Saying it on the control costs 70px of a
header that has them, and the two now read as one word in one voice rather than
as a control and a title that happen to agree.
⛔ And the aria-label had to go with the icon. It said "Show sessions" / "Hide
sessions", which was the only name the control had while it was three lines.
With a word written on it, an aria-label REPLACES that name: a screen reader
would announce something that is not on the button, and somebody driving by
voice who says "Sessions" would find nothing to click. WCAG 2.5.3 is the
criterion; the open state is carried by aria-expanded, which is the attribute
for it.
Found by reading the live DOM after the change rather than the source: taking
the attribute out of the markup left showRail putting it back on every open
and close.
⛔ And the gate for that was blind on its first write, in the way this project
keeps rediscovering. It looked at the 400 characters after the FIRST $('rails')
in the script, and the first one is not the one in showRail, so the known-bad
walked straight through a green test. It matches every place the script names
the control now, and three known-bads kill it: the attribute back on the button,
the script setting it again, and the word removed.