Skill: chef (weekly nutrition-variety advisory)

Use when: the chef-weekly heartbeat fires, or someone asks about their eating variety, plant diversity, or meal monotony.

One-line scope: a weekly, observational variety report built from photo-logged meals. It reports what was eaten. It never reports what anyone is.

What v1 does - the defensible core

1. Plant count. Distinct plant species over the trailing window against a target of 30, per the American Gut Project (McDonald 2018, mSystems, n approximately 10,000: 30+ distinct plants per week associates with greater gut-microbial diversity than 10 or fewer). A COUNT is immune to the portion-estimation error that kills photo-derived micronutrient math - grams do not change whether a leek was eaten.
2. Variety and rotation. Protein and headline-dish frequency, flagging monotony and suggesting rotation. Framed as variety and interest, which is the thing that actually drives adherence.
3. Gap board. Food groups absent from the window, ranked by how easy each is to add. An observation about the log, never a claim about a body.
4. Next week's shape. Three to five concrete, locally available suggestions, one of them a New Plant of the Week - a species not seen in the window.

Hard refusals - encode these, never cross them

• Never diagnose or hint at a deficiency. Not "you may be low in X", not "you could be short on X". That word requires bloodwork.
• Never recommend a supplement, by dose, brand, or in general.
• Never recommend a calorie target, a deficit, or a weight-loss protocol.
• Never construct a restriction, elimination, or fasting protocol.
• Never assert a physiological state from food - not microbiome composition, not inflammation, not micronutrient status.
• Never do micronutrient, vitamin, or deficiency quantification. Deliberately deferred: per-item portion grams are produced by an equal-split heuristic, so any nutrient figure derived from them is a division wearing a lab coat.
• Never frame low-mercury fish repetition as a mercury problem. Salmon at roughly 0.022 ppm is about 7x below the FDA 0.15 ppm Best Choices threshold; cod and shrimp likewise. Their repetition is a VARIETY point only. The mercury frame is legitimate ONLY for genuinely high-mercury species (swordfish, tuna, king mackerel, shark, marlin) recurring three or more times, and even then with the real ppm attached.

These are not stylistic preferences. This skill reads a food diary and talks back about it, which is exactly the shape of a tool that drifts into clinical claims it has no basis for. The refusals are the reason it is safe to run unattended.

Where the pieces live

Piece	Path	Purpose
Gate	scripts/gather-chef-weekly.sh	Fires only on the report day, with enough logged days, once per ISO week
Analyzer	scripts/chef_plant_count.py	Deterministic mapping and counts
Taxonomy	data/plant-taxonomy.json	Free text to canonical species, non-plant list, FDA high-mercury table
Prompt	scheduled-tasks/chef-weekly-prompt.md.template	The heartbeat prompt
State	<state_dir>/chef-weekly-state.json	ISO-week dedup

Extend the plants map as new foods show up. unresolved_tokens in the analyzer output is the candidate list.

Data dependency

This skill reads the meal tables written by the companion food-logging plugin (bfl_meals, bfl_days by default, overridable via meals_table and days_table). It writes nothing back. Without that plugin, or an equivalent schema, there is nothing to count.

Reading the number honestly

Report distinct_plant_count as it comes back. Do not adjust it upward.

The taxonomy has been corrected once already for exactly this: it was counting red onion and green onion as two species, romaine and iceberg and "mixed greens" as three, and "mixed vegetables" as a plant despite naming no species. The result was consecutive weeks reporting above target against a log that showed the target reached once in four months.

An inflated counter is a participation trophy. If the number jumps because something changed in the taxonomy or the parser, say which change caused it rather than letting it read as progress.

Standing footer on every chef report

Observational. Built from photo-logged meals with substantial portion error. Reports what was logged, not nutritional status. Nothing here is medical advice - for anything about deficiency, supplements, or targets, that is a doctor and a blood panel.

Escalation - the only one

If the log shows a genuinely stark pattern, such as a food group absent for 30 or more days, state the observation and suggest raising it at the next routine medical visit. Do not name a nutrient. Do not prescribe a fix.

Deliberately deferred

Do not build these without an explicit decision:

• Micronutrient and vitamin quantification - blocked on per-item portion truth.
• Deficiency inference - permanently out. Requires bloodwork.
• Microbiome composition claims - not measured by a food log.
• Recipe generation and meal scheduling - only after the counts are trusted.