The clothing design studio lets a designer change a garment by talking to it. Earlier today it learned to leave refusals alone ("not black, make it cream"). The one sentence that fix still got wrong was a question: "isn't navy better?" was read as turning navy down.
Questions turned out to be a bigger hole than that one sentence. A keyword reader looks for words anywhere in a clause, so it treats every clause as an order:
- "Is embroidery more expensive?" switched the print to embroidery.
- "Does fleece work with embroidery?" changed both the fabric and the method.
- "Which is cheaper, dtg or screen print?" chose DTG.
People also point back at what they saw a minute ago, and the reader had no memory of it. "Go back to the cream" ran one undo, so if the last change was the print method, the method went back and the garment stayed navy. "Undo the last two", "the previous colour" and "start over" were not understood at all.
Three kinds of question
A new step sorts each clause before the parser sees it:
- A question for information changes nothing. The studio answers it in the log. Cost questions are answered from the designer's own supplier sheet: each alternative the question names is applied to a copy of the design and costed the same way as the costing panel. "Which is cheaper, dtg or screen print?" gives both landed costs and names the cheaper one. Fabric and method questions get the method's note and the things to check with the maker for that pairing.
- A suggestion asked as a question ("isn't navy better?", "what about a crewneck", "why not embroider it", "what if we made it bigger") is tried. Every change is undoable, and someone who asks "what about navy?" wants to see navy. The log marks it "Trying it (you asked)".
- A polite order ("can you make it cream", "could you move it up a bit") is still an order.
Voice input has no question marks, so the sorting uses the words that open the clause. A typed "?" helps: "maybe olive?" is a suggestion, "navy?" on its own is a question.
References to earlier versions
The command box now gets the session's history, oldest version first, and resolves what a sentence points at:
- "Go back to the cream" or "the hoodie looked better" restores that value only, keeps everything done since, and says how many changes ago it was seen.
- "The previous colour", "the original size", "the old position" and "the other one" find the last version where that attribute was different (or the first one) and restore just that attribute.
- "Undo the last two" and "go back three steps" undo that many steps. "Back to the original" and "start over" return to the first version, and one undo brings the current one back.
How it was checked
I wrote 41 sentences while building this, each with the design it should leave behind after a scripted session (a black tee, then cream, a hoodie, a moved and enlarged halo, navy, screen print). The code on main got 17 right. The new code gets 41, which is expected for sentences written alongside it. So after freezing the code I wrote 20 more. main got 12 of those and the new code got
- The miss was "what if we did it in white", read as a question. That pattern was added afterwards,
so those 20 are no longer held out.
The cost answers are checked against the costing function itself: an answer equals the landed cost of the design with that change applied. Like the costing panel, they are only as good as the supplier sheet behind them.
None of these sentences came from a real designer, and browser speech recognition adds its own mistakes. Recorded feedback from a designer working on a real garment is still the test that matters. The script and full results are in the venture's eval/history-questions-2026-11-17.md.