Tools used

search_foods
create_diet_item
update_diet_item
get_diet_item
analyze_training_history
evaluate

Procedure

Daily calorie target = maintenance exactly. No deficit, no surplus. Protein runs high (2.0g x bodyweight kg). The mechanism is slow: muscle gain and fat loss happen simultaneously because the body recovers from training using existing fat stores while synthesizing new tissue from a strong protein supply.

Recomp works for three populations and almost no one else:

Untrained or detrained — recovering muscle is metabolically easy.
Carrying significant body fat (>20% men, >28% women) — fat is readily available substrate.
Returning from a long break — "muscle memory" effects.

If the user is a lean intermediate / advanced lifter, recomp will stall in 4-6 weeks. Surface this at activation.

Decision rules for the coach

Maintenance is the entire prescription. No targeted carb distribution, no carb cycling, no refeeds. Eat at maintenance every day.
Protein is non-negotiable. 2.0g x bodyweight kg minimum. If the user logs < 1.6g/kg for two consecutive days, flag it.
Weekly weigh-in for noise; monthly photo for signal. Weight shouldn't change meaningfully on recomp. Body-composition shifts show in photos / measurements / strength numbers, not the scale. Schedule a monthly check-in card.
Training intensity is the lever. Recomp depends on a productive training programme. If the user has no active training skill, flag at activation and suggest pairing with upper-lower-4day or full-body-3day.
Strength is the proxy. Use analyze_training_history to pull the 4-week trend on main lifts. Strength up + weight flat = recomp is working. Strength flat + weight flat = recomp has stalled, suggest transitioning to a bulk-cut cycle.
Indefinite duration. No length cap. The skill stays active until the user transitions; suggest re-evaluation every 8 weeks.
Inline caveats. Recommending sustained caloric prescriptions touches medical-adjacent surface — emit the caveat per respond-to-medical on any message that prescribes macros.