Skip to content

Update the Rubric

The rubric is not a pass/fail scoring sheet. It is a structured guide that tells the agent how to think through its own evaluation: what to look for, how to weigh different signals, and how to arrive at a classification. The agent reads the rubric and applies it as judgment, not as a mechanical formula.

There are two rubric memory stores, one for each workflow:

Memory Store Used By
Rubrics - Mansa Daily inbound email processing
Rubrics - Mansa Meeting Weekly post-meeting recap and pipeline updates
  • After a team retro where Mansa consistently missed or misclassified a type of email
  • When BKR Capital’s investment thesis shifts (new sectors, new stage focus, new geography)
  • When you want to change how the Meeting Agent evaluates or categorises decisions from Deal Intro meetings
  • When new deal-flow patterns emerge that the current rubric doesn’t handle well

Step 1: Edit the Rubric File in SharePoint

Section titled “Step 1: Edit the Rubric File in SharePoint”

The rubric source files live in the Mansa Rubrics SharePoint folder. Open the file for the rubric you want to change and edit it directly. Rubric files are plain Markdown.

Make your changes and save in SharePoint before touching the Claude Console.

Go to platform.claude.com/dashboard and sign in.

In the left sidebar, expand Managed Agents and click Memory stores.

In the memory stores list, find the store prefixed “Rubrics - “ that corresponds to the workflow you’re updating:

  • Rubrics - Mansa → daily inbound processing
  • Rubrics - Mansa Meeting → weekly meeting recap

Click on the store to open it.

Find the rubric file inside the store and click to open the editor. Select all the existing content and replace it by pasting in the updated text you saved in SharePoint.

Save the file. The updated rubric takes effect on the next agent run.

A rubric file is a Markdown document structured around the agent’s decision-making process. For the daily inbound rubric, typical sections include:

  • What counts as a relevant lead: signals the agent should treat as strong positive indicators (founder present, stage fit, sector fit, deck attached)
  • What is noise: newsletters, service provider pitches, irrelevant cold outreach
  • Edge cases: how to handle ambiguous signals (e.g. a warm intro with no deck, an email in French)
  • How to handle missing information: what to do when a key signal is absent
  • Output expectations: what the agent’s classification response should look like

Write rubric instructions the way you’d brief a thoughtful analyst. Use clear prose, not a rigid checklist.

If a rubric change causes the agent to misclassify emails or make poor meeting decisions, contact Adullam Technologies to revert to a previous version via the Claude Console API.