Scorecards are the foundation of Craft. Get them right and the AI becomes your best coaching tool. Get them wrong and scores feel random. This guide covers how to write scorecard items that score accurately and are easy to coach from.
Use simple plain English instructions
This makes scorecards easier to maintain, debug, and coach from.
Less is more
Describe the behavior you are looking for in the simplest possible terms. The system is smart; just describe what you want. Cut filler and boilerplate. Shorter prompts are easier to reason about and debug when results look wrong.
Skip examples and quotes
Do not add examples or specific quotes unless you actually expect the rep to say them verbatim or close to verbatim. LLMs are next token predictors and pattern match aggressively. If you add quotes or a “language examples” section, the AI will err toward requiring those exact words. Then you end up adding more and more examples to cover edge cases. Describe the behavior instead.
One behavior per step
If a step checks two things, split it into two steps. A single score on a two-part behavior is ambiguous and you cannot coach from it. It’s fine if the behavior is complex, or has a lot of nuance to how to do it, but don’t conflate two completely separate things into one item.
Each item is scored on its own
Each item is scored in its own completely separate AI call. The AI sees only the transcript, the rep name, the company name, the date, and that one item’s name and description. It does not see the other items, section headers, step order, or weights. This has two consequences:
- Never reference one item from another (“if the previous step passed…”). Items have zero influence on each other.
- Step order is not enforced. An “opened with a strong intro” item has no idea it is step 1. If timing matters, say so in the item itself (“early in the conversation”).
Scores run 0 to 5
You do not have to define the scale, but sometimes it helps to steer the AI more predictably. You can list each score from 0 to 5 with a simple sentence describing when to assign it, or add statements like “do not score above 3 unless the rep did X.” All optional; add them to steps where scores are coming back inconsistent.
Phrase every item so that a high score is the good outcome
There is no support for negative or inverted items. “Did the rep interrupt the customer?” scored 5 counts as a good result in the final score. Write “Rep let the customer finish speaking without interrupting” instead.
The item name is part of the prompt
The AI sees the name and the description together. A vague name like “Step 3” wastes signal, and a name that contradicts the description is a conflicting instruction.
The AI only sees the conversation
It gets the transcript plus basic context (rep name, company name). It cannot score anything that happens outside the conversation, like “logged the lead in the CRM” or “called back within an hour.”
Allow N/A
If the behavior is not always relevant, check “Allow N/A” on the step, and add a simple instruction to the prompt describing exactly when the AI should mark it N/A. The checkbox only gives the AI permission; the conditions have to be in your prompt text. N/A is neutral, not a penalty: an N/A item drops out of the score entirely and does not drag it down.
Weights are applied after scoring, not by the AI
Apply a weight above 1 to steps that matter more or have a higher influence on the sales process. For example, price conditioning effectively deserves a higher weight than asking whether to take your shoes off. But do not write “this step is very important” in the prompt; the AI never sees weights and that sentence just adds noise. See Scorecard Weights for the full math.
Minor things
Write like a human. Use as few words as necessary to express what you want, no more, no less. Proofread each prompt to make sure you did not accidentally add conflicting instructions. If you want structured sections in a longer prompt, use markup tags like <criteria></criteria> rather than markdown headers. Not required, but useful in complex prompts.