The My Words voice tab has no persistent record of what the partner changed. The
reveal highlight is transient and the next edit's selection overwrites it; the
Undo button names only the top of the stack; everything else sits behind the
Debug disclosure. A writer who looks away for ten seconds cannot answer "what did
it change?" — so the awareness half of the dual-channel design (hear the
narration, see the edit land) only works if you happened to be watching.
Why this is cheap
The data is already captured. Each undo-stack entry carries
{ target, splices, description }, and a paragraph splice holds both the old and
the new text of the affected range — that is a diff. What's missing is
retention: the stack caps at 10 and discards entries on undo, so a history view
needs to keep popped entries rather than re-derive anything.
describeOp already produces the one-line summary used on the Undo button.
Scope to decide
- Sidebar diff list vs. inline change marks vs. both.
- Whether entries stay after undo (probably yes, marked reverted).
- Whether clicking an entry re-highlights the passage in the document.
- The spoken half: "what did you change?" narrating the last N moves. Worth
designing together with the visible half rather than separately.
Background
Both halves are recorded in the design notes. Grep the docs for:
Deliberately deferred — the section listing this and the other three gaps,
including the substrate note above.Review-out-loud — the open design axis this pairs with (spoken review).Control invariants — the enforced rules in the voice session, one of which
(no silent failure) is the same "don't leave the writer guessing" concern
applied to refusals rather than to changes.
The My Words voice tab has no persistent record of what the partner changed. The
reveal highlight is transient and the next edit's selection overwrites it; the
Undo button names only the top of the stack; everything else sits behind the
Debug disclosure. A writer who looks away for ten seconds cannot answer "what did
it change?" — so the awareness half of the dual-channel design (hear the
narration, see the edit land) only works if you happened to be watching.
Why this is cheap
The data is already captured. Each undo-stack entry carries
{ target, splices, description }, and a paragraph splice holds both the old andthe new text of the affected range — that is a diff. What's missing is
retention: the stack caps at 10 and discards entries on undo, so a history view
needs to keep popped entries rather than re-derive anything.
describeOpalready produces the one-line summary used on the Undo button.Scope to decide
designing together with the visible half rather than separately.
Background
Both halves are recorded in the design notes. Grep the docs for:
Deliberately deferred— the section listing this and the other three gaps,including the substrate note above.
Review-out-loud— the open design axis this pairs with (spoken review).Control invariants— the enforced rules in the voice session, one of which(no silent failure) is the same "don't leave the writer guessing" concern
applied to refusals rather than to changes.