How ModelMonkey Thinks

As ModelMonkey works on your request, you may notice status cards appearing in the conversation. These indicate that ModelMonkey is engaging in deeper reasoning:

Exploring... - ModelMonkey is analyzing your spreadsheet's structure
Planning... - ModelMonkey is breaking down a complex task into steps
Verifying... - ModelMonkey is checking its work for accuracy

Each card shows a pulsing indicator while active, then a checkmark when complete (or an error icon if something went wrong). You can click on any card to expand it and see what ModelMonkey is thinking about.

ModelMonkey uses specialized reasoning modes (internally called "subagents") to handle complex tasks more effectively. Each mode is optimized for a specific type of thinking:

Explore analyzes your spreadsheet to understand its structure before making changes. It identifies:

• Key elements like tables, headers, and data ranges
• Naming conventions and formatting patterns
• Relationships between sheets
• The overall purpose of your spreadsheet

Plan breaks down complex requests into ordered, actionable steps. It considers:

• What needs to happen first, second, third
• Dependencies between operations
• The most efficient approach to achieve your goal

Verifier audits completed work for quality. It checks:

• Completeness - was everything requested actually done?
• Formatting consistency - do new cells match your existing style?
• Formula accuracy - are calculations correct?
• Data integrity - do totals reconcile?
• Hardcoding - are there values that should be formulas?

ModelMonkey decides when to use these reasoning modes based on the complexity of your request:

Explore is used when:

• First working with an unfamiliar spreadsheet
• The task requires understanding the spreadsheet's structure
• You ask questions about how data is organized

Plan is used when:

• A task requires 3 or more coordinated changes
• Multiple sheets or ranges are involved
• The order of operations matters

Verifier is used when:

• Complex formulas have been written
• Bulk data changes were made
• Financial calculations or critical data is involved

For simple, single-cell operations, ModelMonkey works directly without invoking these modes.

While ModelMonkey decides autonomously when to use these reasoning modes, your prompts can influence that decision. Here are examples that often trigger each mode:

To encourage exploration:

• "Analyze this spreadsheet"
• "What's the structure of this data?"
• "How are these sheets related?"
• "Explain what this spreadsheet does"

To encourage planning:

• "Break this into steps before starting"
• "Show me your plan first"
• "This is complex - plan it out"
• "Walk me through how you'll do this"

To encourage verification:

• "Verify your work"
• "Check if the formulas are correct"
• "Audit what you just did"
• "Make sure the totals add up"
• "Double-check for errors"

Example: After ModelMonkey creates a financial summary, you might say:

"Verify that all the formulas are correct and the totals reconcile"

ModelMonkey will then run a verification check and report any issues found.

When ModelMonkey verifies work, it produces a structured report with findings across several categories:

Completeness - Was the requested task fully implemented?
Formatting - Do new cells match existing formatting conventions?
Accuracy - Are formulas correct and logically sound?
Balances - Do aggregate calculations reconcile?
Hardcoding - Are there values that should be formulas instead?

Each category shows either a checkmark (passed) or specific issues found.

Overall Verdict:

• Pass - All checks passed, work is ready
• Pass with Warnings - Minor issues that may need attention
• Fail - Significant issues that should be addressed

Sample Verification Report:

[COMPLETENESS] All requested changes implemented
[FORMATTING] Warning: Row 15 uses different date format (MM/DD vs DD/MM)
[ACCURACY] All formulas verified
[BALANCES] All totals reconcile
[HARDCODING] Warning: Cell D12 contains hardcoded value, expected formula

Overall: PASS WITH WARNINGS

When verification identifies issues, you have several options:

For warnings: Review each one and decide if action is needed. Some may be intentional (e.g., a hardcoded value that's meant to be a constant).

For failures: Address these before considering the work complete. You can ask ModelMonkey to fix them:

• "Fix the formatting issues found"
• "Convert that hardcoded value to a formula"
• "Address the verification warnings"

For false positives: Occasionally verification may flag something that's actually correct. Use your judgment and domain knowledge.

Verification is a helpful safety net, but it's not infallible. Always apply your own judgment for critical work.