Search: fifteen ideas

As she types, she sees what kinds of things match and how many, before running anything. “del” already tells her delivery is a tag, a section and a signal, and that it comes up 64 times. The layout follows Photos: an icon for the kind of thing, the typed letters in full-strength text with the rest of the word lighter, and a count on the right.

She can type things that aren't words and still get a precise answer: a participant code, a timecode, a feeling in her own language. The search recognises what she probably means and offers it one row below plain text. Pressing ↩ still does the plain search. It never quietly decides “p3” means a person.

Once a suggestion becomes a token, she can change its meaning without retyping: quotes p2 said, or quotes that mention p2 (when the moderator or another participant talks about them), or everyone except p2. A tag token switches between tagged with it and containing its words. These distinctions matter in analysis, and here each is one click. It is the down-arrow menu on Mail's tokens.

One page answers “where did this come up?” across every lens. The groups always appear in the same order, so she learns where to look. Each group has a way into its home lens that carries the search with it: “Show all 23 in Quotes →” opens Quotes already filtered. The scope bar says plainly what was searched.

The search ends in analysis, not just a list. She finds every mention of “assembly”, ticks the ones that count, and tags them all at once. Tags already on each hit are visible, so she can see what's coded, and hits already carrying the tag she's applying are shown faded. MAXQDA and ATLAS.ti users expect this; research repositories mostly don't offer it.

Inside a transcript, search is find: nothing disappears, and ⌘G steps from hit to hit. The playhead jumps to each hit, so she can hear every mention of “delivery” in s2 in about a minute, including tone, hesitation and laughter, which the text doesn't show. Marks on the scrubber show where the talk about it clusters.

Each hit shows the sentence it matched in and why. The reason may be a different form of the word, text she edited herself, a tag definition, or the moderator's words. She can judge a hit without opening it, and a result that doesn't visibly contain her word still makes sense.

The moderator often introduces the topic word: “tell me about delivery”. Counted with everything else, the interviewer's questions look like evidence. Participants' own speech comes first; moderator turns are folded into their own group and not counted as findings, but they're still one click away when she wants the question that prompted an answer.

When nothing matches, she gets a next step. There's a “did you mean” built from words actually in the study, then ways to widen the search: hidden quotes, the full transcripts, the rest of the folder. Each shows its count before she clicks, so none of them is a guess.

She can see patterns across studies for the same client: does delivery cost come up in the returns pilot too? Every hit starts with the name of the project it came from. People appear only as speaker codes, because a participant agreed to one study, not to being linked across several. The last line says which projects were not searched and why, so a missing project never looks like a missing finding.

Clicking into an empty field brings back this project's recent searches, tokens included, so yesterday's question is one keystroke away. Recents belong to the project: nothing from another client's study ever shows up here. They can be cleared, and they're forgotten when the app quits unless she saves one (idea 15).

Only numbers that won't change get quoted in a readout. While the study is still being indexed, counts read “12+” and the menu says so at the bottom, the way Photos does. When everything has been searched the “+” goes and the count is exact. A number never goes down as more results arrive, so she learns to trust it.

When she types a question rather than a keyword, the menu also offers a written answer with citations, from the chat lens. It's clearly labelled as a Claude step, and it only runs when she picks it. ↩ still does the ordinary keyword search, so a half-typed question can't start a paid call by accident.

The search belongs to the question, not the page. With “p2 · hidden costs” in the field, switching lens shows the same question from a new angle: p2's quotes, p2's sessions, the cells of the heatmap p2 contributes to. She never retypes it, and each lens has just one search field that works the same everywhere.

A question she keeps asking becomes a place. “Pricing pain: hidden costs or price surprise, with frustration” is saved into the project, appears in the sidebar and stays up to date as coding changes, like a Smart Mailbox in Mail. Saving is a deliberate act, so ordinary searches still reset between sessions.