part one · observation

Observation

Going to watch the person work, not asking.

Observation is the act of being in the room while the person does the activity. Not in a focus group. Not in an interview room. Not over a screenshare summarising what they would normally do. In the room, in real time, while the activity happens.

This is the discipline most teams skip and most teams pay for. The reason is simple: observation is uncomfortable. The person being watched is uncomfortable. The watcher feels intrusive. The schedule is harder to arrange than a 30-minute call. So teams substitute interview, and a brief is built on what people say they do — which is reliably different from what they actually do.

Three kinds of input, ranked

Rank	Method	What you get	What you miss
1	Observation in the field	The actual activity, the actual environment, the actual interruptions, the actual workarounds	Things they only do once a quarter
2	Recorded session of the activity	The activity, the workarounds, much of the friction	Environmental context, interruptions
3	Structured interview about the activity	A described version of the activity	Workarounds the person has stopped noticing, environmental friction, real timing
4	Survey	A summary of summaries	Almost everything

The corpus pattern: observe first. Use interview to fill in details. Use survey only to quantify a pattern that observation has already named.

What you watch for

Three classes of signal.

1. Workarounds. The places where the person has built their own bridge over a gap the system left. Workarounds are the most valuable Discovery signal — they are evidence of pain that has already been measured by the person paying it.
2. Friction. The pauses, the re-entries, the let me find that, the alt-tabs, the post-its. Friction is rarely complained about because the person has internalised it.
3. Domain language. The actual words the person uses for the things in their work. They are almost never the words on the screen.

The note-taking method matters less than the discipline of writing what was seen, not what was concluded.

The interview that earns its place

Interviews follow observation, not precede it. They are not for asking what is hard about your job. They are for asking what was that thing you did at 09:14, and why?

A good post-observation interview is short. Half an hour. Five questions, all anchored to specific moments the watcher saw.

• I noticed at 09:14 you opened the spreadsheet, edited a name, and closed it. What was that?
• At 09:31 you went to the wall calendar before answering a question. What were you checking?
• At 09:48 you switched to the second monitor for about thirty seconds. What were you reading?

Specific moments, named in time. The person's answer is the most honest answer the chain will ever get from them, because it is anchored to something concrete and recent.

When observation is impossible

Sometimes it is. Regulated environments, sensitive client contexts, fully remote-async workflows. The corpus's fallback, in order:

1. Recorded session with the person's permission.
2. Reconstructed session — they walk through their actual work using their actual tools while screensharing, narrating in real time. Worse than observation because the narration changes the timing, but better than interview alone.
3. Pair-and-shadow — a colleague of the person sits with them and reports back. Adds noise but holds the witnessed discipline more honestly than interview.

What does not substitute: a survey, a stakeholder description of the person's work, or anyone else's interpretation of the person's life.

What gets recorded

The output of observation is a small artifact: the observation note. Lives in the project space, attached to the initiative brief.

Observed:        Gal, exam grader, Wed 2026-04-22, 08:50–10:30
Activity:        Grading a batch of five Computer Science final exams
Watcher:         Alex (PO), Maya (Designer)

Timestamps:
08:53  Opens the LMS, navigates: Courses → CS101 → Submissions →
       Filter unread. Two clicks were wrong direction. (Friction: nav.)
08:57  Pulls up the rubric on second monitor. Switches between the
       student PDF and the rubric repeatedly.
09:14  Edits a student's display name in a spreadsheet, then returns.
       (Workaround: LMS does not support Hebrew names properly.)
...

Domain language used by Gal (verbatim):
  "the answers" (she means student submissions)
  "the boxes" (she means rubric criteria)
  "send back" (she means require resubmission, not return for editing)

Stopped noticing:
  Gal stops looking at the timer in the LMS after the first 10 min.
  She has internalised that it is not accurate.

The note is not a polished document. It is raw. It is what the brief is later built on.

Part 2 — Person & Moment →