Math via Code - Skill Setup¶

Source: tommulkins/hermes-skill-math-via-code Stars: 0 ⭐ | License: MIT Created: June 24, 2026 Category: Development / Accuracy / Financial Analysis

1. What It Is¶

A Hermes Agent skill that enforces a hard rule: all arithmetic with 3+ numbers goes through code execution. No exceptions. Born from repeated errors in financial modeling - SDE calculations, P&L analysis, and facility rollups where a single transposed digit cascaded into incorrect deal evaluations.

Core philosophy: LLM arithmetic accuracy degrades sharply past ~150k context tokens and is unreliable even below that. Code is always right; in-head math is not.

What It Handles¶

Financial analysis: SDE, P&L, margins, royalties, annualizations
Growth rates, ratios, percentages, weighted averages
Unit conversions, projections, multi-facility rollups
Reading data from files/APIs and computing derived values
Excel-specific pitfalls (whitespace stripping, data_only=True, etc.)

When NOT to Use¶

Single-operation arithmetic on two numbers (5 * 3, 100 - 20) - these are safe in-head
Pure text manipulation or formatting tasks

2. Prerequisites¶

Hermes Agent
No additional packages required (uses built-in execute_code)

3. Installation¶

hermes skills install https://raw.githubusercontent.com/tommulkins/hermes-skill-math-via-code/main/SKILL.md --category software-development --name math-via-code -y

Verify¶

hermes skills list | grep math-via-code

4. How It Works¶

Decision Flow¶

Numbers provided? → Count them:
  ├── 0-2 numbers → Safe for in-head math (e.g., 5 * 3)
  └── 3+ numbers → ROUTE TO CODE (execute_code or Python script)

Procedure¶

Determine inputs - Capture numbers as variables from conversation, files, or APIs. Never re-type from memory.
Write the calculation - Use execute_code for in-memory math; use terminal with project venv for packages (openpyxl, pandas).
Self-check - Include assertions or cross-checks against source data.

Example: SDE Calculation¶

# Instead of in-head math, the skill routes to:
revenue = 327158
direct_costs = 192193
gp = revenue - direct_costs

overhead_items = [960, 1351.92, 9000, 1440, 1137.24]
overhead_royalties = round(revenue * 0.07, 2)
total_overhead = sum(overhead_items) + overhead_royalties

sde = gp - total_overhead
print(f"SDE: ${sde:,.2f}")

5. Data Sources¶

Source	Guidance
Conversation	Capture numbers as Python variables directly
Excel/CSV	Use `openpyxl` with `data_only=True`; strip whitespace from all cell values
API responses	Parse JSON, validate numeric fields, handle null/missing values
Multi-file	Cross-reference with assertions (e.g., `assert sum(details) == total, "Mismatch"`)

6. CorpusIQ Use Cases¶

Use Case	Application
Lead qualification math	LTV calculations, churn projections, CAC payback periods
Pricing analysis	Margin calculations, tier comparisons, bundling economics
Growth metrics	MoM/ YoY growth rates, cohort retention math
Agent cost tracking	Token cost aggregation, per-agent cost allocation, budget burn rates
Financial reporting	Revenue rollups across multiple data sources

7. Troubleshooting¶

Symptom	Cause	Fix
Skill doesn't trigger	Calculation is simple (2 numbers)	Only 3+ number operations trigger the skill
Excel values wrong	`data_only=False` returns formulas	Use `openpyxl.load_workbook(file, data_only=True)`
Whitespace breaks numbers	Cells have trailing spaces	`.strip()` all cell values before conversion
Assertion fails	Source data inconsistent	Report the mismatch - don't silently proceed

← Skills Catalog | June 24 Evening Discovery →

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects