Weights, deal breakers, and the scoring formula

The prompt format

Each criterion carries an explicit weight out of 100, and deal breakers are stated separately as reject rules rather than as heavily negative weights. Weights and rejections do different jobs: a weight moves a score, a rejection ends the conversation.

Score this company as a prospect from 0 to 10.

[weight 30/100] Company size fit
Headcount between 20 and 400. Score full marks in the middle of that range, partial marks at the edges, zero outside it.

[weight 25/100] Market complexity
Sells into more than one country, or into a market where no single database has good coverage. Evidence: the countries column and the market notes column.

[weight 25/100] Timing signal
Currently hiring for a sales, growth or revenue operations role, or has opened a new location in the last twelve months. Evidence: the hiring signal column.

[weight 20/100] Commercial maturity
Has at least one named person with a commercial title. Evidence: the decision maker column.

REJECT IF
- The company is an agency reselling lead data or list building to end clients.
- Headcount is under 5.
- There is no website and no other online presence.
- The company operates only as a marketplace with no direct sales motion.

OUTPUT
A score from 0 to 10, then a two sentence rationale naming which criteria carried the score and which evidence you used. If a criterion has no evidence in the row, say so rather than assuming.

Four rules for the weights

1

2

3

4

1. 1LLM score is the mode this course teaches. Company ML and People ML sit beside it and need training data first.
2. 2The weight marker, written into the prompt itself, right after the criterion it belongs to.
3. 3The same weight as a control. Change either and the other follows, so the prompt is always the source of truth.
4. 4Total Weight, adding to 100 out of 100. The product tracks this for you, which is why the rule below is not arbitrary.

• Weights add to 100. If they do not, you are not being explicit about what matters most.
• Four to six criteria. Fewer and the score is coarse, more and everything averages to the middle.
• No single criterion above 40. If one thing really decides it, that is a filter, not a scoring criterion.
• Every criterion names the column that provides its evidence. This is what makes the rationale checkable.

1. 1

   Open the Toolbox and pick the Scoring Agent

   It is under Custom research, not in a category of its own. Configure it as a column on your table.

   

   1

   1. 1Custom research, in the Toolbox sidebar. Scoring Agent is in here.

2. 2

   Paste the weighted prompt

   Keep the weight markers and the reject block exactly as formatted. The structure is what the agent reads.

3. 3

   Run it on 20 rows first

   It costs nothing, so there is no reason not to. Read all 20 rationales.

4. 4

   Check the spread

   If every row lands between 5 and 7 your criteria are not discriminating. If everything is 0 or 10 your reject rules are doing all the work. You want a spread across the range.

   

5. 5

   Fix the criterion, not the score

   When a row scores wrongly, read its rationale and find which criterion caused it. Edit that criterion. Never patch a score by adding a rule about a specific company.

6. 6

   Run it across the table

   Still free. Run it on everything, including rows you think will fail, because the failures are what confirm the model works.