How to use class recordings without rewatching everything

Why full rewatching feels productive and isn't

Watching a lecture a second time feels like learning because everything looks familiar. Familiarity is not the same as being able to do it. In data science, the test is whether you can write the pandas step or explain why a model overfits with the recording closed. So the recording is only useful at the moment you find you cannot — which means you need to try first.

Keep a timestamp log

During a live session, or within a day of it, keep a plain text log with one line per topic. Mark how confident you felt on a 1–3 scale.

# Session 14 - Model evaluation (hypothetical example)
00:04  recap of train/test split                     3
00:21  confusion matrix built by hand                2
00:48  precision vs recall, fraud example            1
01:15  ROC curve and AUC                             1
01:52  break
02:05  cross-validation with KFold                   2
02:40  doubt: why stratify on imbalanced data?       1
02:55  assignment brief                              3

It takes a minute per topic. The confidence number is what drives everything else. A 3 means you could teach it; a 1 means you followed it while it was on screen but could not reproduce it.

If you attend live, note the time a doubt was asked — including other learners' doubts. The answers to other people's questions are often the most useful minutes of a recording.

The recall-first rule

For every line rated 1 or 2, before opening the recording:

1. 1Close the video and your notes.
2. 2Write down, or type in a blank notebook, what you remember: the idea, the code, the example.
3. 3Compare with the class notebook or slides.
4. 4Only where your attempt was wrong or blank, open the recording at that timestamp.

Step 2 is uncomfortable, and it is what makes step 4 short. You arrive at the recording knowing exactly which minute you need.

Turn the log into a revision queue

If you keep logs for several sessions, a few lines of Python sort them into the order you should revise: lowest confidence first, oldest first among equals.

log = """\
14 00:21 confusion matrix built by hand 2
14 00:48 precision vs recall, fraud example 1
14 01:15 ROC curve and AUC 1
14 02:05 cross-validation with KFold 2
15 00:10 class weights for imbalance 1
15 01:30 threshold tuning on validation set 2
15 02:20 calibration curves 3
"""

queue = []
for line in log.strip().splitlines():
    parts = line.split()
    session, stamp, confidence = int(parts[0]), parts[1], int(parts[-1])
    topic = " ".join(parts[2:-1])
    if confidence < 3:
        queue.append((confidence, session, stamp, topic))

queue.sort()
for confidence, session, stamp, topic in queue:
    print(f"[{confidence}] session {session} at {stamp}  {topic}")

assert queue[0] == (1, 14, "00:48", "precision vs recall, fraud example")
assert len(queue) == 6

Each week, work from the top of that list for whatever time you have, and raise the confidence number after a successful redo. When everything is a 3, the recording has done its job.

When a full rewatch is justified

There are a few cases where watching a whole segment again is the right call:

• You missed the session entirely. Watch it once in full at normal speed, keeping the log as you go, then apply the recall-first rule.
• The topic was the foundation for everything after it, such as a first session on joins or on train/validation/test splits.
• You rated most of the session a 1. That is a sign to ask for help as well as rewatch; send your mentor the timestamps.

Speed and attention

Playback at 1.25× or 1.5× is fine for review of material you partly know. For a first viewing of a missed class with new code, normal speed is usually better, and pausing to type the code yourself matters more than the speed setting. Keep the class notebook open beside the video and run each cell when the instructor does.

Avoid watching recordings on a phone for coding topics. You cannot type along, and without typing along, little stays. Conceptual segments — an explanation of bias and variance, a discussion of a case — work reasonably on a phone.

A weekly routine for a two-session week

What to ask a mentor with timestamps

Timestamps make questions far easier to answer. Compare "I didn't understand ROC curves" with "At 01:15 in session 14, why does moving the threshold trace the curve instead of changing the AUC?" The second gets a precise reply in minutes. Our guide to asking a doubt your mentor can answer in one reply goes further.

Recordings are most useful when they back up a live class rather than replace it. The Data Science programme records its live online sessions so missed or difficult parts can be revisited, and if you are still choosing a format, the recorded versus live self-assessment may help.