Ask what to train today

Morning, undecided. The routine has options; recovery is mixed; you have 45 minutes; the knee's not 100%. You ask the agent to read the state and propose something coherent — not a fixed prescription, a starting point you can accept or push back on.

What the agent needs to hear

Behind your sentence, the agent works backwards to one piece: whatever constrains today. Time available, body state, equipment access, mood — the constraint is the lever that turns a generic suggestion into a useful one. Without it, the agent falls back to your active routine and recent recovery state and proposes from there.

Constraints narrow the read. « 45 minutes » caps the duration. « Knee’s stiff » steers away from lower-body push. « Hotel gym, dumbbells only » filters equipment. Naming the one that matters most today is enough — « what should I train today? » with no extras works fine; you just get a wider proposal.

What the suggestion reads from

Four inputs feed the proposal. The active routine sets the baseline — if today’s pull day is queued, that’s the starting frame. Recovery readiness reads which muscle groups have been worked recently, so the agent doesn’t propose back after yesterday’s row session. The constraints you named — time, equipment, body limits — narrow the queue. The progression plan picks the next intensity for the relevant lifts, so the suggested weights match where you actually are, not last month’s numbers.

The output is one suggested session with its queue, target sets, and a one-line rationale: « push day, since you trained pull yesterday and shoulders are fresh ». The rationale is the load-bearing part — it tells you which inputs drove the pick.

The suggestion is a starting point, not a prescription

The proposal is a Bayesian best guess, not a rule. The agent reads the signals, weighs them, and offers one coherent session. You accept, adjust, or override — all three are fine. Asking and then ignoring is also fine; the suggestion gets logged either way, so the system learns what you actually do vs. what it proposes.

The rationale matters more than the verdict. A suggestion without a why is hard to argue with. A suggestion with a why lets you push back on one input: « I trained pull yesterday but my back’s fine, swap to pull again ».

When the agent gets it wrong

Three failure modes show up. The agent ignored a constraint — « you said 45 minutes, this is 75, trim it » — and you talk it back down. The agent missed a recent signal — « knee was stiff this morning, push not pull » — and you re-anchor the body state. The agent suggested a session you’ve already done this week — « I trained back yesterday, pick a different day » — and the queue rebuilds.

In each case you keep talking, you don’t edit a form. Name the input the agent missed and the suggestion regenerates around it.

What makes the suggestion worth asking for

Three things decide whether the suggestion helps your day: the constraints you give are the ones that actually matter today (time and body state weigh heaviest — equipment third), the rationale reads as a thing you can argue with (a suggestion without a why is hard to override; with a why you can swap one input), and the suggestion is treated as a starting point, not a verdict (the warm-up tells you the truth — adjust the session in flight if the body says different). Ask, accept or override, repeat — over weeks the system learns your preferences from the gap between what it proposes and what you do.