The Generative AI Pack

Describes version 4.70.1

Local AI analysis for your studies: session summaries, a cited findings report, and grounded question-answering, all running entirely on your machine. Nothing about your study leaves your network.

Installing the Generative AI Pack

AI Insights, Findings and Export Report

The analysis is three entries in the left menu, each holding one kind of work.

✨ AI Insights is where you decide things. It has five tabs across the top of the page. Status says what this study's data will and will not support, before you generate anything: whether sessions have notes, whether task spans are timed, whether outcomes are scored. It is first on purpose, because that is worth knowing before a thin report tells you the same thing slowly. ✨ Generate findings report is on this tab too, under that reading, because this is where you learn whether generating is worth doing. While it runs, its progress takes the button's place. Proposed is what the sessions suggest, one card each. Dismissed is what you turned down, with Restore on each. Summary is the generated narrative itself, and you can edit it. Ask is your questions and their answers.

Findings lists the findings you have kept, in your order, and opens on the first one. New finding starts one of your own.

Export Report is one screen, with two documents beside it: the detailed report as an HTML package or a PDF (the study's full record: the synopsis, the export's own details, and as applicable the findings, task data, survey data, use errors, highlight clips, participant information and task materials) and the executive summary as a .pptx deck (a cover slide, a summary of the findings, one slide per finding, and a slide carrying the export's details).

Status is how the report stands: when it was last generated, whether any insight is still undecided, and what the file will contain section by section, each marked as generated, computed or yours. Almost every section is worked out when the file is written and so cannot be out of date. Exactly one is written by the model, and that one can be: it is marked out of date when a finding or a recommendation has changed since it was written. ✨ Generate findings report appears in that row, and only when generating would change something, so a button offering to spend several minutes rewriting the same words is not on the screen. While it runs, its progress takes the button's place, with a Stop beside it. Nothing asks you to confirm, so export first if you want to keep a copy of the report you have.

Summary: the generated text, in your words if you prefer

The Summary tab on AI Insights holds the narrative the model wrote, and it is an editor rather than a reading: change a word, a paragraph, or all of it. What is in that editor is what goes into the detailed report and the deck, so this is the one place to look if you want to know exactly what generated text your documents will carry.

Your version and the model's are kept separately. What the model wrote is never destroyed by your editing, so the study can always say which words came from the machine and which from a person, which matters if anyone ever asks. Clearing the editor completely hands the text back to the model rather than leaving an empty summary.

Do not overwrite this on generation, above the editor, protects what you have written. Generating a report normally produces a fresh narrative; with this checked, your text carries over to the new report instead and the model's new wording sits underneath it. Leave it unchecked and the next report shows the model's words again. Either way the earlier version stays in the study.

It is a per-report switch, not a study setting: you are protecting the passage you rewrote, not declaring a rule for every report this study will ever produce.

The report is assembled, not written. Its counts, medians and citations are computed from your study and its quotes are checked against what was recorded; one opening paragraph is generated. It proposes nothing: everything on it is settled by the time you open it, so reading it is reading rather than deciding. A report covers:

• What we tested · How it went overall · What got in people's way · Task by task · What we recommend.
• Sections that have no data say so rather than filling themselves in. Where no task was timed or scored, How it went overall and Task by task say that completion was not measured and describe the sessions in words instead; where a task was defined but nobody attempted it, the report says that about the task and nothing more. You will see this on imported sessions, which arrive with no task spans until you mark them.
• Every claim carries a citation naming the participant and the moment, like ▶ Part02 @ 6:35 "…": click it to jump to that exact moment in the recording, even in another session. Review opens at that click, not before. A session with no participant assigned is cited by its session number instead, S2 for the second recording.
• A collapsed Researcher appendix carries the exact methodology detail - themes, severity, per-task scores - computed by the app, never paraphrased by the model.

The summaries are prepared ahead of time. As soon as a study has sessions with notes in them, the app summarizes each one in the background, so the activity strip at the foot of the window may readthe activity line in the top bar may read AI Analysis: preparing session 2 before you have asked for anything, and the findings report screen says the same. That is groundwork, not the report: you can still generate, and a report you ask for simply runs after the summaries it needs. Stop preparing session summaries, on that screen, or Stop in the activity strip's detail list, leaves that session for the report to summarize later; nothing is lost but the head start.

Findings: keep what the sessions suggest

On AI Insights → Proposed, each thing the sessions suggest is a card: a count the app computed from the sessions it cites, such as 8 of 10 participants, the claim, a quote, the heuristic, and three buttons. Keep makes it a finding, and it joins the Findings list in your order. Edit keeps it and opens it. Dismiss tells the next report not to propose it again and moves it to the Dismissed tab, where Restore brings it back. Nothing is called a finding until you accept it, which is why the tab is called Proposed. Under Seen once are moments one participant hit that look like an issue on their own; keep one if the issue is plain without corroboration. Generating again proposes only what is new: it is told what you kept and dismissed.

A finding has three tabs of its own. Details carries its name, where it came from, its frequency and your description. Supporting Evidence carries the moments it rests on. Recommendations carries what it calls for.

A finding is a claim, a count and its evidence. Select one in the rail to open it. The title and the claim are yours to write, and they save as you type. The count is computed from the supporting evidence, distinct participants over a denominator you choose: all participants, or the participants who attempted one task. The default follows the evidence: the task's attempters when all the supporting evidence sits in one task, otherwise all participants. It is never sorted by and never typed.

Evidence is a note, a speech line, a use error or a task outcome from a session. Each row names the participant it came from and what was said or seen, with everything you can do to it underneath: Cut highlight here, Remove, Watch, its moment, and a supports | refutes switch with the current side pressed. Add evidence… picks more from a session, with a search. One row is marked ★ Finding highlight: that is where the finding's highlight is cut, there is one per finding, and it is the only row without a Cut highlight here button, because it already is the highlight. Double-click evidence to watch it in Review. Drag a finding up or down the list to change your order (Alt+↑ and Alt+↓ do the same from the keyboard)Use ▲ and ▼ to move a finding up or down the list (⌥+↑ and ⌥+↓ do the same from the keyboard), and Delete in the rail removes it; deleting a kept proposal lets the next report propose it again, which is what Dismiss prevents.

The highlight is a short video of the finding for the deck. Keeping a proposal cuts it for you at the evidence marked ★, from the session's video as Review shows it, in the session's current layout. How long it runs depends on the evidence. A participant quote or a task attempt has a start and an end, so the highlight runs from two seconds before it starts to three seconds after it ends, never shorter than ten seconds nor longer than forty-five. A note or a use error has a moment rather than a span, so the highlight runs from five seconds before it to fifteen after. Transcripts indexed before this build have no ends recorded: their quotes get the moment window until the media is indexed again (Index Media, on the Review page). It is saved in the study's Highlights folder and listed on the Highlights tab in Review, captioned Finding: … with the finding's title, so it is a highlight like any other. To cut it somewhere else, choose Cut highlight here on a different row: that marks it ★ Finding highlight, cuts a new highlight at its moment and removes the old one. On the ★ Finding highlight line under the list, Play opens it, Adjust… opens Review on that session with the highlight's start and end loaded in the marking controls so you can move them and save, and Cut discards those edges and cuts the original span over. Removing the evidence marked ★ removes the highlight. A finding you add yourself gets its highlight when its first supporting evidence is added, which marks it. The highlight is a starting point: the moment worth showing rarely sits exactly where the note was typed, which is what Adjust… is for.

✨ Suggest evidence…, beside Add evidence…, asks the model to look for evidence you have not cited. The app gathers what could be cited from every session (what participants said, what observers noted, use errors, task outcomes), ranks it by the finding's own words and shows the model the most likely candidates; the model picks the ones that plainly bear on the finding and says why, in a sentence. They appear under the evidence, each with the model's reason and Watch, Add and Skip; Watch plays the recording right there over the same span the highlight would use, and pauses at its end, so you can check it says what the model thinks before adding it (every piece of cited evidence has Watch too); nothing is cited until you add it, and the model can only point at things that exist in the sessions. It runs in the background like a report; the activity strip says suggesting evidence for a finding. A finding you add yourself is the usual case: you noticed something the report did not, and the evidence is the part still to gather.

Design recommendations from this finding, under the highlight, lists what the finding calls for. New recommendation… records one on the Design Recommendations page, linked to this finding and citing its supporting evidence, and opens it there to write. ✨ Suggest recommendations… asks the model for two to four, drawn from the finding's claim, its heuristic and its evidence and told what the study already records; each has Keep, which records it the same way, and Dismiss. A kept suggestion is a draft to edit on that page, never a conclusion. Open…, or a double-click, goes to a recorded recommendation.

Add my own finding…, above the list, is for a finding you noticed that the report did not. It starts empty, with Add evidence… and Suggest evidence… ready; it carries no ✨ because nothing about the finding itself is generated.

The deck. The executive summary, on the Export Report page, saves the kept findings as a PowerPoint file: a cover, a summary of the findings with their counts, one slide per finding in your order, and a closing slide on how the findings were made. A finding's slide carries its count in the wording you chose, the claim, the evidence marked ★ as a quote with who said it and when, the recommendations recorded from the finding, and the highlight itself, embedded and playable with a click. The deck uses PowerPoint's own layouts and placeholders throughout, with the theme's fonts and colors rather than fixed ones, so your company template applies cleanly: in PowerPoint choose Design, then Browse for Themes, pick your template, and the deck takes its fonts, colors and positions. Long text shrinks to fit its box. Highlights need the study's highlight files where the study is; a finding whose highlight has not been cut gets a text slide.

Ask AI

✨ Ask AI, on the Ask tab of AI Insights, shows a question box. Type a specific question and press Enter or Ask: "Where did people struggle with checkout?", and answers it from only your study's data, with clickable citations. The question joins the top of the Answers list straight away, selected, showing its progress until the answer replaces it, and every question you have asked stays there to reread. Each answer opens with its question, so copying an answer into a report or slide brings the question along. If the data doesn't answer the question, it says so rather than guessing. Ask AI is read-only: it never changes the study.

Make this a finding, above a finished answer, turns it into one. An answer and a proposed insight are the same material, a claim with checked quotes under it, so the finding carries the answer's quotes as citations on the moments they came from. It opens on the Findings page with its title selected, because it arrives named with your question and only you know what the finding is called. The answer stays where it is.

What needs generating again

Generating a report is the slowest thing the Generative AI Pack does, so it is worth knowing what actually requires it.

• Nothing you do to a finding needs it. Keeping, editing, reordering, adding evidence, writing a recommendation and cutting a highlight all belong to you, not to the report, and the exports read them as they are at the moment you export.
• Exports never need it. Results by task, the charts, survey results and every other figure in the detailed report are computed from your study when the file is written, so they are current by construction. Export again and the numbers are current; there is nothing to regenerate first.
• Changing the study's data does need it if you want the report's prose to reflect the change. New sessions, newly marked task spans, scored outcomes and renamed tasks all change what the report would say.
• Transcription and on-screen text indexing never need repeating. They read the recording, and nothing you do afterward changes what was said. This is the expensive part, and it is safe.

Session summaries look after themselves. Each one remembers the session data it was written from, so marking task spans or renaming a task means the sessions affected are summarized again the next time you generate a report or ask a question, and the rest are reused. You do not need to ask for this and there is nothing to clear.

AI history

Answers and session summaries are kept. The Ask tab lists every question you have asked, so you can find last week's answer; a new answer joins the top of its list when it finishes. The findings report is different: a study has one, and generating again replaces it. A report is a draft from the sessions at one moment, and after the data changes the old draft is not worth rereading; the HTML… and PDF… exports are how a copy is kept. What lasts is what you keep from it: the findings, which are yours and are never replaced by generating again.

There is no command that throws all of it away at once, and none is needed: generating again replaces the report, dismissing handles an insight you do not want, and the delete control in the rail handles a single answer. Session summaries are never removed by any of these, which is the point - they are the expensive part and are still true.