Full-service restaurant prep cook - knife-spec card SOP - for every dish, what cut, how long, how thick, how many pieces per portion; with cards built right a trainee holds the station in three days, and without them whoever cuts decides what the dish tastes like

You are the head chef's assistant at a full-service restaurant. Write the "knife-spec card" building process as an archivable standards document for the prep station. Context: full service, many dishes and many specs, prep staff rotate often, and flavor drift usually starts on the board rather than at the range. The card program has one goal: a different person cuts, the same dish comes out.

1. Card fields (fixed template)
- Dish / main ingredient / cut (julienne / slice / cube / baton / segment / mince)
- Dimensions (length x width x thickness, or the cube-size basis)
- Pieces per portion OR weight per portion (one or the other, matching the cost card's basis)
- Aromatics ratio (scallion / ginger / garlic / chili, shape and quantity)
- Prep stage (cut raw / cut after blanching / cut after marinating)
- Vessel (plating and staging containers)
- A "why this cut" line (even cooking / coating adhesion / shape retention / speed - the real reason; this line is never omitted)

2. Building process (order is not negotiable)
1. Scope: build the first [20] cards for high-frequency, many-hands dishes - do not chase the whole menu; a card's value concentrates where hands rotate
2. Set the standard live: the head chef cuts the reference sample by hand and measures on the spot (calipers or a comparison object for size, a digital scale for weight) - every number comes from measurement. You (the AI) only transcribe spoken specs into structured card text: invent one number and the card is void
3. Photograph: the reference sample from the front, from the side, and next to a hand or chopsticks for scale
4. Write the card: fill the template; for dimensions add a lay description (e.g., "3 mm, about a coin") - the description is a reference aid; judgment always follows calipers and scale
5. Post: cards go up at the prep station (or onto the tablet), grouped by menu family or firing station
6. Version log: any spec change must record who / when / why (dish upgrade / ingredient change / cost adjustment) - an unlogged change counts as no change, and the next audit against the old card will surface two competing "standards"

3. Acceptance basis (whether cards are working)
- Pull a cook who did not build the card, have them cut from the card, and set the result beside the reference sample: three or more mismatches (any of dimensions / piece count / aromatics counts as one) means retrain, not replace
- Re-cut and re-compare [3] random cards monthly, photographed side by side

Red lines (own section):
- Spec numbers are never copied from other menus, cookbooks, or "industry practice" - this house's knives, ingredients, and ranges decide this house's specs
- No card, no menu listing - build the card before the dish launches; off-menu specials get their spec set verbally on the spot and back-filled into a card after service
- Inventing a number voids the card (addressed to both the AI and any human running the process: transcribe, never complete)

Output format: sectioned, archivable document; include one blank card template (complete fields, printable) and one comparison sheet template (photos side by side, item-by-item ticks).

Why it is written this way

When a dish drifts, the reflex is to blame the range - but most flavor drift in full service starts on the board: for one dish, the shreds two cooks cut differ by a factor of two in thickness, and coating, evenness, and cook time all move with it. No line cook, however steady, can cook flat what the board cut uneven. That is why the card's success criterion reads "a different person cuts, the same dish comes out": cards exist for handover, and dishes that never change hands do not need one - hence scoping the first batch to high-frequency, many-hands dishes.

"Invent one number and the card is void" addresses two audiences: the human running the process and the AI helping. The only legitimate source of a spec number is the head chef's reference sample measured live - calipers and scale decide. The most dangerous failure of an AI-assisted card build is helpful completion: a blank thickness field gets filled from general knowledge, that number becomes the house standard, and nobody remembers it was invented. Transcription without completion is the AI's entire job boundary here.

The "why this cut" line is what separates a card from a checklist. A rushed trainee does not consult cards, but does remember cause-and-effect like "thick shreds lose the coating" - knowing why tells them what can never bend and what can flex with circumstances (swap the cut of meat, and the person who understands the principle knows what to adjust). Cards without that line train photocopiers, not prep cooks.

The version log guards against the two-standards incident: the chef verbally changes a spec, the card stays, and three months later an audit against the card leaves the cutter indignant and the checker certain - both holding a "standard". Logging makes the change itself visible, which is the least a card needs to remain the single standard. And acceptance is designed as "pull someone who did not build it and compare cuts" rather than "check the cards are posted" - a card's value is proven at the board, not on the wall.

Variables