Prompts for planning training with AI (and why they work)

1.The difference that changes everything else

In a chat with no access to your data, the prompt has to carry everything: who the athlete is, what they did, what equipment exists, how many days they train. That is why the training "mega-prompts" people share run to fifteen lines — they are compensating for missing information with text.

With a connector attached, half of those lines are waste: the assistant can go and look. Your job becomes stating the decision — the goal, the constraint, the date — and asking it to check what it wrote. Shorter prompts, better results.

2.The prompts worth keeping

To open the day: "What needs my attention today?" — it comes back with who has no week planned, whose week is empty and whose block is ending with nothing after it, across every active athlete at once.

To create: "Create a 5-week hypertrophy block for Sofia, 4 days, upper/lower, starting Monday. Hold the sets and add load week to week; week 5 is a deload." Note that it states the progression — without that you get five identical weeks.

To adjust: "Sofia’s back volume is under target this week. Add two sets on a day with room, without touching anything else." And to review: "Compare week 1 with week 4 of this block and tell me where the load went up and where it stalled."

3.Why these work

Three properties, all of them copyable. The first is naming the athlete: with no name the assistant has to guess or ask, and either way you have spent a turn. The second is giving the real constraint — days, equipment, the joint that hurts. A plan that is perfect for a gym that is not yours is no use.

The third is asking for the check in the same sentence. "Create the block, then show me the volume per muscle" is worth far more than "create the block", because the assistant can count what it has just written. It is the difference between trusting and checking, and it costs nothing.

4.The prompts that look good and are not

"Make me the best possible training plan" has no answer. Best for what, for whom, with what equipment, across how many days? The assistant will invent the missing answers, and you will get a generic plan that reads as a specific one.

"Use the latest science" is worse, because it sounds rigorous. An assistant has no live access to the literature and does not know what came out this month; what it does is write in the voice of someone who does. If you want defensible ranges, ask for ranges and check them — the ones we use are in the sets-per-week guide, with the sources at the foot.

And "don’t forget anything" is not an instruction. Say what cannot be missing: a warm-up, core work, a day with nothing that loads the knee. What is implicit to you is not implicit to it.

5.The prompts for weekly upkeep

Creating blocks is what gets demonstrated; upkeep is what eats the week. To log what arrived by message: “Sofia did Monday’s session — bench 4×8 at 52.5, row 4×10 at 45, pull-ups 3×6. Log it.” One sentence, a whole session, without opening the record.

For the week after: “Copy Sofia’s week 3 into week 4 and add 2.5 kg on the compounds.” And for what usually goes unnoticed: “Which exercises has Sofia not moved up on in three weeks?” — that is the question that catches a stall before it turns into a drop-off.

None of these needs you to remember anyone’s state. That is the difference between a connected assistant and a notebook: you can ask instead of having to know.

6.When the result is not what you wanted

The wrong reaction is to start over. If the block came back with too much volume, do not ask for another block — say what is wrong: “You dropped the shoulder work and put twenty-two sets of chest in. Fix both and leave the rest as it is.”

Correcting beats restarting for two reasons. The first is that the assistant already has the conversation’s context and will not repeat the same mistake. The second is that starting over returns a plan that differs in five other things you had already approved.

And when the correction fails too, the problem is usually the original request, not the assistant. Go back, write the constraint that was missing, and ask again — stated, not implied.

Tools

Guides

Frequently asked questions