GoldLift 的 AI 教練如何根據訓練紀錄、漸進超負荷與恢復安排下一組

Disclosure: I'm the developer of GoldLift. Everything below describes the app's own code as of October 2026.

Quick answer: GoldLift's coach does not invent weights with a language model. The AI writes your programme and talks to you, but the weight and reps for every set come from a fixed, deterministic rule that reads your last session at that exercise: reps hit, RPE, whether you collapsed or ground out a rep, which equipment you use and whether you are in a deload week. Recovery is tracked separately with a fatigue score built from your own logs. The coach sees that score in every conversation and can propose a deload week, which you confirm before anything in your plan changes.

This post exists because the fair question about any "AI coach" is whether it actually adapts to you, or just hands out a template. Here is what GoldLift does, in enough detail to check.

Two layers: the AI talks, the rules count

There are two parts with different jobs.

• The language model builds your first programme from your goal, level, equipment and training days, answers questions in the chat, explains decisions and can propose changes (a swap, a deload week).
• The progression rule is plain code. It decides the number in every set cell, the "ghost" target you see before you lift. Every screen uses the same rule, so the chat, the workout screen and the suggestions never disagree.

That split is deliberate. A language model is good at explaining why; it is not reliable at arithmetic on your history. So the numbers never come from it.

The progression rule, in priority order

For each exercise, the rule looks at your previous session's working sets (warm-ups are ignored) and picks one outcome:

The coach puts the reason in one line under the exercise, for example: "Last time 80 kg × 10 @ RPE 7.5. Today 82.5 kg."

Details that stop it from being a generic template

Rep ranges depend on the goal of the set. A target of 6 or fewer reps gets a range of +2, 7 to 11 gets +3, 12 or more gets +5. You can cap it with your own "Max reps" setting.

Weights move by what your equipment can actually do. Barbells move by your plate step (2.5 kg by default, adjustable). Dumbbells and kettlebells move along the real rack, so it never asks for a 13.5 kg dumbbell. When the next dumbbell is a big jump (8 to 10 kg is +25%), it first asks for more reps at the current weight, so you do not land on the heavier one and fail.

Straight sets move together. The weight goes up when your first set reached the top of the range and the later sets got close, with an allowance for fatigue: one rep less on set two, two reps less from set three on. Without that, the fifth set of a 5 × 3-5 would keep you on one weight for months.

It reads your sets together. If you did 8, 8 and then 14 reps at the same weight, the first set clearly had more in it, and the rule treats it that way. One easy set (RPE 6.5 or lower) also tells it the sets before it at the same weight were no harder.

Big misses are corrected in one go. When you are far outside the range, for example after guessing a starting weight, it estimates your one-rep max from the set (Epley formula) and jumps to the right load. That jump is capped at +15% up or -20% down, so one bad day is not over-corrected.

Pyramids and top sets are recognised. If you worked up to a heavy top set with back-off sets, only the top set carries the progression, and the lighter sets keep their ratio to it.

You can switch it off. Progressive overload is a setting. When it is off, the cells simply mirror last session.

Recovery: a fatigue score from your own logs

Separately from the per-set rule, GoldLift keeps a fatigue score from 0 to 100. It is built only from what you log, never from a fixed "deload every 4 weeks":

• the fatigue rating you give after a session (1 = fresh, 5 = exhausted), and whether it is climbing week over week;
• the average RPE of your sets this week;
• how many days in a row you trained and how many sessions in the last 7 days;
• the acute-to-chronic workload ratio, meaning this week's volume compared with your own 4-week average;
• whether your estimated one-rep max on your main lifts dropped compared with the two weeks before;
• the weeks since your last deload or naturally light week.

Rest days pay fatigue back, up to 30 points. Above 50 the risk is "high", above 70 "critical".

The coach sees it. Your fatigue score and overtraining risk go into every chat with Rex. Above 70 the coach is told to lean toward recovery advice, and with an elevated risk it does not recommend more volume before addressing recovery.

Deloads. Ask "do I need a deload?" or tell Rex you feel beaten up, and it can propose a deload week: fewer sets and roughly 10 to 20% lighter loads for 7 days. The recommendation also weighs the weeks since your last light week (6 by default). You see a card and decide; nothing changes until you tap it. During that week the progression rule switches off completely. Every set is prescribed at the deload load and never heavier than last time.

What the coach knows when you chat

When you ask Rex something, it is not answering from a blank page. Every message includes:

• your sessions from the last 8 weeks, with strength trends on your main lifts (a lift needs at least 3 sessions over 2 weeks before it gets a trend, so it does not over-read one good day);
• sets per muscle group over the last 4 weeks, compared with what your goal needs;
• your food logs from the last 2 weeks and your weight trend;
• the workout you are doing right now, set by set, even if it has not been saved yet.

When there is not enough data, the coach is told that a number is missing, not that it is zero. So a new user does not get told they are "under-training legs" after one session.

Honest limits

• The automatic deload analysis needs at least 10 logged sessions. Before that, there is too little history to judge.
• The rule is only as good as what you log. Without RPE it falls back to rep-based double progression, which works but is slower to react. On the last set of an exercise the coach asks for an RPE when you have not given one.
• The programme itself is written by a language model from your onboarding answers. It is personal, but a first plan, not a coach who has watched you lift. The per-set rule is what makes it adapt from week two on.
• GoldLift is young, so there are few independent reviews yet. This post explains the mechanism so you can judge it yourself.

FAQ

Does GoldLift's AI coach use my training history?
Yes. Every set's target is computed from your last session at that exercise (reps, RPE, equipment), and the chat coach sees your last 8 weeks of sessions, strength trends, volume per muscle, food logs and weight trend.

Does GoldLift do progressive overload automatically?
Yes. It adds reps until the top of your rep range, then adds weight by your equipment's real step. It holds after a hard set at RPE 9 or higher and drops the weight after a collapse. You can switch progressive overload off in the settings.

Does GoldLift take recovery into account?
Yes. A fatigue score from your session ratings, RPE, consecutive training days, workload versus your 4-week average and strength trend goes into every conversation with the coach, which can propose a deload week. During a deload, progression is switched off.

Are GoldLift's weights generated by ChatGPT?
No. The language model writes the plan and the chat answers. The weight and reps per set come from a fixed rule in code, so they are consistent and explainable.

Is GoldLift a good free alternative to Fitbod?
For lifters who follow a structured plan, yes: it prescribes weights per set from your history, handles deloads and lets you ask the coach why. Fitbod's algorithm has more years of data behind it, and Fitbod costs about $96 a year after its trial; GoldLift is free.

GoldLift on the App Store: https://apps.apple.com/be/app/goldlift/id6758995677. Website: https://www.goldlift.app.