Analyze Usability Sessions at Scale With Clip Evidence

Workflow area: Research & Synthesis · Tier: Core · Hands-on time: ~35 minutes

The old economics of usability analysis were brutal: an hour of session tape cost two hours of analysis, so teams either ran fewer sessions or skimmed the tapes and trusted memory. AI analysis flips the cost — extraction is nearly free — but it introduces a new failure mode: an issues list that reads confidently and cites moments that do not say what the summary claims. The clip link is your contract with the truth. An issue without a clip is an anecdote; an issue with an unwatched clip is a liability.

When to use — and when not to

• Use it when you have five or more recorded sessionsbelow that, watching the tapes yourself is faster and richer than configuring extraction.
• Use it when findings will drive a prioritized fix listseverity plus clip evidence is exactly the format engineering and PM can act on without attending the sessions.
• Use it to make stakeholders watch the two minutes that mattera clip of a real user failing beats any slide; extraction finds the clip, you deploy it.
• Skip it for sessions where the signal is emotional, not behavioraltone, trust and discomfort live between the words; transcripts flatten them and the model summarizes the flattening.
• Skip the automation if your transcripts are garbageheavy accents, crosstalk or bad audio produce transcript errors that become confident false issues; fix transcription quality first.

The recipe

1. 1.

   Get clean, timestamped transcripts with speaker labels. Maze, Dovetail and Condens all auto-transcribe uploads. Spot-check one transcript against two minutes of video: if the words are wrong, stop and re-transcribe — everything downstream inherits transcript quality.

2. 2.

   Extract issues with a schema that forces evidence. Run each session (or batch, if the tool supports it) through this prompt — in Dovetail or Condens use their AI analysis with the same instructions, or paste transcripts into Claude:

   Copy

   You are analyzing a usability session transcript for UX issues.
   Product context: [ONE PARAGRAPH — what was tested, the tasks given]
   Transcript with timestamps: [PASTE]
   
   Extract every usability issue you can support with evidence.
   For each issue output:
   - issue: one sentence, phrased as observed behavior, not opinion
   - task: which task it occurred in
   - evidence_quote: verbatim participant words or described action
   - timestamp: start time of the moment
   - severity_draft: 1 (cosmetic) to 4 (task-blocking), with one line
     of reasoning tied to task impact
   Rules: no issue without a verbatim quote and timestamp. Do not merge
   distinct issues. Do not infer feelings the participant did not state.
   If a moment is ambiguous, list it under "needs human review" instead.
   

3. 3.

   Merge across sessions into an issue table. Feed the per-session outputs back with: "Merge issues that describe the same underlying problem. Keep every timestamp as separate evidence. Count sessions affected." The table you want looks like:

   Copy

   ISSUE                                 SESSIONS  SEV  EVIDENCE
   Users miss the save confirmation      5 of 12    3   S2 04:12, S5 11:40,
   and re-submit the form                               S7 02:55, S9 08:31,
                                                        S11 06:02
   Filter labels read as buttons         3 of 12    2   S1 09:18, S4 03:44,
                                                        S8 12:09
   

4. 4.

   Watch every clip before the issue graduates. This is the step that cannot be delegated. Open each timestamp and watch 30–60 seconds around it. You are verifying three things: the moment happened, it means what the issue says, and the severity is right. Expect casualties — a "task-blocking" issue that turns out to be the participant misreading the task brief dies here. Log each clip as confirmed, reframed (issue is real, description was wrong) or rejected.

5. 5.

   Set final severity yourself and ship the fix list. Severity is a judgment call about user impact and frequency — the model's draft is a starting point, not a verdict. Publish the table with clip links inline so any skeptic can watch the evidence in one click. The fix list that survives is short, watchable and very hard to argue with.

Make it yours

• Dovetail highlights as the clip currency: create a highlight per confirmed timestamp and tag by issue — your issue table becomes a living reel you can replay in any prioritization meeting.
• Maze for task-level metrics plus clips: pair the extracted issues with Maze's misclick and completion data; an issue confirmed by both a clip and a metric jumps the priority queue.
• Claude for the cross-study memory: keep a Project with each study's confirmed issue table; before every new study, ask "which of these issues have we seen before?" — recurring confirmed issues are your strongest roadmap ammunition.

Quality check

• Every issue on the final list has a watched clip"confirmed" means a human watched it, not that a timestamp exists.
• Issues describe behavior, not diagnosis"users re-submit the form" is evidence; "the button lacks affordance" is your interpretation and belongs in the recommendation column.
• Session counts are honestan issue seen once is a data point, not a pattern; the table says "1 of 12" out loud.
• Rejected extractions were logged, not deletedthe rejection rate is your calibration data for how much to trust the next extraction run.
• Severity survived your reviewif you accepted every draft severity untouched, you graded nothing.