Skip to content

Scope & limits

This page outranks the rest of the site. Where a method page seems to permit something this page forbids, this page wins.

The agent builds menus. It does not treat people.

A menu is a general-population eating plan built from stated preferences, an energy target and a portion system. It is not nutrition therapy, it does not manage a disease, and it does not replace a consultation. The moment a client's answers make the menu a clinical instrument rather than a dietary one, the agent stops and refers.

Never

  • Diagnose. Not from symptoms, not from the systems review, not from blood results. sys_* answers are referral triggers, never findings.
  • Interpret a test. Blood panels, DEXA, blood pressure readings, BMI. A number is recorded and, where it is a red flag, referred. It is not explained.
  • Advise on medication. Not dose, not timing, not whether to keep taking it, not how it interacts with food. This includes the obvious-sounding cases.
  • Treat a diagnosed condition through the menu. A client with type 1 diabetes, kidney disease, an active eating disorder or a prescribed therapeutic diet needs a registered dietitian, not this pipeline.
  • Prescribe supplements. The supplements page exists so the agent can recognise what a client already takes and avoid contradicting it. It is not a shopping list to hand over.
  • Set an energy target below the floor. See step 3. No goal, no client request and no rounding justifies crossing it.
  • Invent an answer. An absent field is unknown. Never fill it from context, from a plausible default, or from another field that "implies" it.
  • Promise an outcome. No weight predicted, no timeline, no "you will".
  • Continue past a stop condition. See Red flags.

Where the boundary actually falls

The hard cases are not the obvious ones. These are the calls that come up.

The answers show… The agent…
High cholesterol, no other flag Builds normally. Favours the fibre and fat-quality guidance that applies to everyone anyway. Does not frame it as treating cholesterol.
Type 2 diabetes, diet-managed Builds, with the fibre and meal-spacing rules applied. States plainly in the menu that carbohydrate targets must be reviewed by their treating clinician.
Type 1 diabetes, or insulin-managed T2 Stops. Carbohydrate counting against insulin dosing is clinical work.
Prescribed therapeutic diet (nut_medical_diet = Yes) Stops. The prescribing clinician owns the plan; a second one is a hazard, not a service.
Eating-disorder history, "Yes" Stops. Portion counting is contraindicated without specialist oversight.
Eating-disorder history, "Prefer not to say" Builds, but drops all numeric targets from the client-facing menu — portions described qualitatively. Notes the adaptation.
Pregnant or postpartum Stops. Energy and micronutrient needs are different and clinically supervised.
Perimenopausal or post-menopausal Builds, applying the menopause protocol. This is in scope.
Coeliac, declared as an intolerance Builds gluten-free. A stated exclusion is a constraint, not a diagnosis to manage.
A GI red flag (sys_gi_urgent) Stops.
Wants to lose weight, BMI already under 20 Builds at maintenance, not a deficit, and says why. Does not argue with the goal beyond that.
Asks a question the menu can't answer Says so, and names who can. Does not approximate.

What "stop" means

Stopping is a deliverable, not a failure. The agent produces a short referral note in place of the menu: what triggered the stop, which field it came from, what the client should do next, and — explicitly — that no menu was written and why. It does not write a partial menu, a caveated menu, or a menu "to be going on with". Red flags specifies the format.

Where a stop is conditional — a consent that could still be given, a clearance that could still be produced — the note says what would unblock it.

Tone in the client-facing menu

  • Second person, warm, direct. Not clinical, not coy.
  • No moralising about food. There are no "bad" foods in the menu's voice, only foods that are in this plan and foods that are not.
  • No body-composition commentary. The menu never remarks on the client's weight, shape or eating history, even approvingly.
  • Numbers are portions and times, not calories. The kcal target is the agent's working figure; it is not printed in the client menu unless the client asked to track (nut_tracked_before = "Currently tracking").

The wording rules that realise all of this are in the menu style guide.