Important Links

• Starting template for task prompt

• Prompt Generator - If the Prompt Generator is used, the task must introduce added complexity. Any needed information can be  added or adjusted. 
  Examples: 

  • Tax complexities (e.g. eligibility + pathway selection) 

  • Financial complexities (e.g. NPV/IRR/payback that won’t contradict) 

    • Feel free to be creative with adding complexity and different assumptions to the prompts to increase complexity. 

• Realm platform (log in using expert.micro1.ai email through Google)

• Example task prompt

• Example Golden Rubrics spreadsheet

Instruction Guide

Purpose

This guide explains how to create one complete Arden task: a realistic clean-energy project, a tax-credit optimization question, and the rubrics used to evaluate model answers and reasoning.

You define the project → the model responds → we score correctness and reasoning. The general scope of this will be to: 

1. Define the 20 required inputs (Location, Technology, Financial).

2. Write a prompt using those inputs.

3. Validate and pass Energy Engineer review.

4. Write Response Rubric (20+ criteria, plus negatives).

5. Chain of Thought Rubrics (20+ criteria, plus negatives).

6. Test 4 models — all must score <60%.

7. Submit.

By the end of this guide, you will:

• Design a technically plausible clean-energy project (PV + optional storage, heat pump, EVSE, etc.).

• Build a scenario-specific tax optimization problem (ITC vs PTC, bonuses, credit strategy).

• Create a Golden Response Rubric and a Golden Chain-of-Thought (CoT) Rubric — these are your ways of encoding the correct answer and logic for the task.

• Run initial tests on Realm with the assistance of Rhea to ensure the task is challenging for current models.

Sources and Resources: 

You also have access to a Resource List (IRS/Treasury guidance, maps, NREL tools, etc.). Use it as your primary research base and as the source library for the “source” + “quote” columns in your rubrics. You are welcome to use other sources as you see fit -- your research process is a part of your expertise.

Adding Complexity:

 Wherever you see a note saying “Good place to add complexity,” that is an explicit invitation to use your tax and structuring expertise to make the scenario harder for current models, while still answerable by a careful human.

Below is a brief index of all the steps, followed by additional detail for each. Not all the steps occur independently -- for example, you will likely develop the prompt as you develop the inputs, and you will develop the rubrics as you add complexity to the prompt. Same with adjustments when you make changes. But the time estimates will give you an approximate breakdown of how much attention each component requires.

Step #1: Confirm Assignment, Create a Task on Realm, and Pick a Building (est. 15-30 mins depending on research complexity)

1.1 Create a new task on Realm (name as [State Abbreviation]-[Project Type]-### based on your assignment)

1.2 Check your assignment - state & project type

1.3 Choose a realistic subject property

1.4 Assume and process special location status

Step #2: Fill All 20 Mandatory Inputs (est. 30 mins-1 hr making complexity-increasing component decisions while working on prompt)

2.1 Location Inputs (5)

2.2 Technology Inputs (8)

2.3 Financial Inputs (7)

Step #3: Write the Narrative Task Prompt (est. 2 hrs, overlap with previous section)

3.1 Narrative

3.2 Task Request (the “prompt asks” your rubrics will grade)

3.3 Legend

3.4 Validate with Rhea and submit for Review

Note on Core Evaluation Types & Rubrics

Step #4: Incorporate Energy Engineer Feedback (est. 15-45 mins depending on project soundness -- make sure everything is adjusted)

Step #5: Build the Response Rubric (est. 2 hrs)

5.1 What the Response Rubric Scores (by prompt ask)

5.2 Table Structure and Sources

5.3 Coverage Guidelines

Step #6: Build the CoT Rubric (est. 2 hrs, overlap with above)

6.1 What the CoT Rubric Scores (by prompt ask)

6.2 Table Structure and Sources

6.3 CoT Coverage Guidelines

Step #7: Run Model Tests and Score With Help of Rhea (est. 2-3 hrs to generate and go through four external model responses and score each using rubrics)

Step #8: Final Submission and Handoff (est. 30 mins)


You can proceed to the first steps here!