Blocked or Random? How to Structure the Session AI Gave You

Ask ChatGPT, Gemini or any of the swing-analysis apps to build you a practice plan and you will get something that looks superb on paper. Twenty balls with the 7-iron working on shallowing the club. Fifteen with the 5-iron. Ten drivers, focus on staying behind the ball. Finish with wedges. Neat blocks, clear intentions, sensible progression.

It is also close to the worst possible way to spend your range time if you want to score better on Saturday.

That is a strong claim, so here is the evidence behind it. The blocked practice structure AI hands you produces very fast improvement inside the session and very poor transfer out of it. You leave the bay striking it beautifully, you arrive on the first tee, and the swing you paid for has gone missing. This isn’t a mystery, and it isn’t your nerves. It is a predictable consequence of how the plan was organised, and it is fixable with about ninety seconds of restructuring.

Why every AI defaults to blocks

The models aren’t being lazy. They are pattern-matching against the enormous corpus of golf instruction written since the 1960s, and that corpus is overwhelmingly blocked. Every drill article, every “ten balls with this feel” YouTube description, every coaching note transcribed onto a forum: blocks all the way down. When you ask an LLM for a practice plan, it produces the median of that literature.

There is a second reason, and it is subtler. AI plans are built around fixing a fault. You upload footage, the tool says your early extension is costing you strike quality, and the plan that follows is engineered to reduce early extension. Repetition is genuinely the fastest route to changing a movement pattern. So the model is correct about the mechanism and wrong about the goal, because your goal was never “change this movement” but “shoot fewer strokes”.

Motor learning research has been consistent on this since Shea and Morgan’s 1979 work, replicated dozens of times: blocked practice wins during acquisition, random practice wins on retention and transfer tests, and the gap is not small. Typical findings show blocked groups performing 30 to 40 percent better in the session itself, then 20 to 30 percent worse when tested a day or two later in a novel context. A golf course is a novel context. Every shot on it is a different club, a different lie, a different target, with no second attempt.

What the tools are actually good at

Before restructuring anything, you need the diagnosis to be right. Here is what genuinely works on phone footage and budget monitors, from someone who has fed a lot of bad video to a lot of confident software.

Swing capture apps. Sportsbox 3D Golf and HackMotion-adjacent tools do real biomechanical work from a single camera, but they are strict about setup. Sportsbox wants 240fps, the phone at hip height, roughly 8 to 10 feet away, and it will reject or silently mangle a clip shot at 30fps from a trolley. If your iPhone is recording in standard mode, you are giving it a quarter of the frames it needs and the club will teleport through impact.

Generalist LLMs on video. Gemini and Claude can both read swing video now, and they are decent at describing large-scale sequencing: “your hips and shoulders are rotating together in the downswing”, “the club is above the plane line at P5”. They are poor at anything requiring frame-level precision. Do not trust an LLM to tell you your shaft lean at impact. Do trust it to tell you your finish position is off-balance towards your toes, because that is visible across many frames.

Budget launch monitors. A Garmin R10 or Rapsodo MLM2PRO gives you carry numbers that are directionally useful and spin numbers that are estimated. The R10 in particular models spin rather than measuring it, which means its spin figures are a function of club speed, ball speed and launch. Use carry distance and dispersion. Treat spin as a rough guide, and never build a shaft-change decision on it.

The useful workflow: capture down-the-line and face-on at 240fps, get one specific fault named by a tool you trust, then structure the session yourself. The pillar piece on turning analysis into a practice plan covers the diagnosis half of this in more depth. What follows is what to do with the diagnosis once you have it.

The split: 60/40, and the second number matters more

Here is the golf practice session structure that fixes the transfer problem. Take whatever the AI gave you, keep the first 60 percent, and rebuild the last 40 percent as randomised, consequence-bearing shots.

A worked example. Say you are an 14-handicapper, your tool of choice has identified an over-the-top move producing a 4 degree out-to-in path with the 7-iron, and the AI plan reads:

AI-GENERATED PLAN (60 balls, 45 min)
Block 1: 15 x 7i, half swings, feel the club drop inside
Block 2: 15 x 7i, three-quarter, same feel
Block 3: 15 x 7i, full swing
Block 4: 15 x driver, apply the feel

Restructure it like this:

RESTRUCTURED (60 balls, 45 min)
BLOCKED PHASE — 36 balls, ~25 min
  12 x 7i half swings, club drops inside, no target
  12 x 7i three-quarter, no target
  12 x 7i full, alignment stick outside the ball to police the path

RANDOM PHASE — 24 balls, ~20 min
  Play Royal Birkdale 1-9 in your head, one ball per shot.
  Change club every shot. New target every shot.
  Full pre-shot routine on all 24. Score yourself.

The blocked phase is where the pattern changes. Twelve balls at each stage is enough; the thirteenth to twentieth balls in a block are mostly you grooving the same rep with declining attention. The random phase is where the pattern gets stress-tested against the thing that actually breaks it, which is not knowing what is coming next.

Note what changes in the random phase: club, target, and shot shape intent, all at once. That is the whole trick. Interleaving works because it forces you to reconstruct the motor plan from scratch on every shot, which is expensive and feels worse, and which is exactly what the first tee demands.

Scoring the random phase so it means something

Twenty-four shots with no scoring is just a warm-down. Give it a number and it becomes a measurement you can track.

Use a simple 3-2-1-0 per shot: 3 for the intended shape finishing within your target window, 2 for a good strike missing the window, 1 for a playable miss, 0 for anything you would reload. Twenty-four shots gives you a ceiling of 72.

Rough benchmarks from tracking this with club players:

The number that matters isn’t the absolute score, it’s the trend, and specifically whether your random-phase score is climbing while your blocked-phase feel stays intact. If your blocked shots are pure and your random score is flat over six weeks, the fault you are working on probably isn’t the one costing you shots. Go back to the footage.

Where the AI plans still earn their keep

Two places, and they’re worth being precise about.

First: short game. A putting or chipping plan is far less damaged by blocking, because the mechanical variation between a 6ft putt and a 9ft putt is small and the skill being built is calibration rather than pattern change. If Gemini gives you a blocked putting ladder, run it as written. The 3-6-9-12 foot ladder, repeated until you hole two in a row from each, is a legitimately good drill and nobody needs to interleave it.

Second: the first two weeks of a genuine swing change. If you are rebuilding a grip or taking a genuinely new takeaway path, you need concentrated repetition before random practice has anything to test. Fourteen days of near-pure blocked work, then introduce the split. Past that point, staying blocked is how people end up with a beautiful range swing and a 19 handicap.

The bit nobody does, which takes four minutes

Log three things after every session in your phone’s notes app: the blocked-phase fault you worked on, the random-phase score out of 72, and one sentence on what broke first when the shots became random.

That third item is the gold. Players discover within about five sessions that the thing that collapses under random conditions is rarely the thing the AI diagnosed. It’s usually tempo, or alignment, or a pre-shot routine that falls apart when the club changes. One 12-handicapper I worked through this with spent six weeks on path work, logged “rushed the transition on every long iron” five sessions running, and dropped four shots off his handicap by fixing his breathing before the shot rather than his club path.

The models will keep giving you blocks because that is what the training data says practice looks like. You now have a structure that keeps the useful part of the block and bolts on the part that makes it survive contact with a golf course. Next time the plan comes back with four tidy blocks, delete the last one and go play nine holes in your head instead.