The IDFA Skill

In Lesson 8, you retrofitted a legacy coordinate-based model to IDFA compliance — converting formulas one by one, validating outputs at each step. Now you have the complete methodology: three layers, four guardrails, naming conventions, the delegated calculation workflow, and the retrofitting process. All of that knowledge lives in your head. The IDFA plugin makes it live in every agent you use.

A SKILL.md file is a structured document that follows the agentskills.io open standard. The Panaversity team has packaged the complete IDFA methodology as a Claude Code plugin — idfa-financial-architect — available from the agentfactory-business-plugins catalog, so you install the skill with two commands and it auto-activates on every financial modelling task. The agent reads it at the start of every session. It does not need to be prompted. It does not need to be reminded. The skill becomes part of how the agent thinks — and when someone mentions a financial model, a spreadsheet formula, or a model audit, the agent applies the full IDFA methodology automatically.

What the Plugin Contains

The IDFA plugin has this structure:

idfa-financial-architect/
├── .claude-plugin/
│   └── plugin.json            ← Plugin metadata (name, version, author)
├── skills/
│   ├── financial-architect/   ← The methodology
│   │   ├── SKILL.md           ← Behavioural guidance, four guardrails
│   │   └── references/
│   │       ├── IDFA-reference.md   ← Enterprise governance, sector naming
│   │       └── IDFA-whitepaper.md  ← Research foundation
│   └── idfa-ops/              ← The operations — the agent's "hands"
│       ├── SKILL.md           ← "Use these scripts for Named Range operations"
│       └── scripts/
│           ├── idfa_ops.py    ← Write, read, inspect, formula, create-range
│           ├── idfa_audit.py  ← Compliance auditor
│           └── recalc_bridge.py ← LibreOffice recalculation
├── README.md
└── LICENSE                    ← Proprietary

The plugin includes two skills working together. The financial-architect skill encodes every concept you learned in Lessons 1 through 8 — the methodology, the guardrails, the naming conventions. The idfa-ops skill gives the agent its "hands" — the scripts that actually read from, write to, and audit Excel models programmatically. When the methodology skill says "write the assumption to the model," the operations skill provides the tool to do it.

Here is how the methodology skill opens:

---
name: idfa-financial-architect
description: >-
  Apply the Intent-Driven Financial Architecture (IDFA) when building,
  auditing, retrofitting, or analysing Excel financial models. Activate when
  the user mentions: financial model, spreadsheet, Excel formula, named ranges,
  cell references, formula tracing, model audit, COGS, revenue projection,
  gross profit, EBITDA, DCF, LBO, comps, three-statement model, budget,
  forecast, variance analysis, what-if analysis, scenario modelling, goal
  seeking, Monte Carlo simulation, model review, or model handover.
  Also activate when the user says "the model is a black box",
  "I inherited this model", "I need to audit this spreadsheet",
  or any similar phrase indicating confusion about how a financial
  model works. Do NOT activate for general accounting questions,
  tax advice, investment recommendations, or tasks unrelated to
  the structure and logic of financial spreadsheets.
license: Proprietary
metadata:
  author: Panaversity
  research-lead: Zia Khan
  version: "1.0"
---

The YAML frontmatter tells the agent when to activate. The description field lists every trigger phrase — financial model, model audit, what-if analysis, "I inherited this model" — that causes the agent to load the full IDFA methodology. Below the frontmatter, the SKILL.md contains the Core Principle, the Three Layers, all Four Guardrails with compliance tests, the Naming Conventions, the Worked Example, the Agent Decision Table, Common Mistakes, and Trigger Phrase mappings. It is the complete methodology in a machine-readable format.

The references/ directory contains the extended reference guide — enterprise governance standards, the five capability tests, retrofitting guidance, complex formula reference (WACC, NPV, Terminal Value, IRR), and sector-specific naming extensions. The agent loads this when the main SKILL.md is insufficient for an advanced task.

Installation Paths

Cowork (Primary Path)

If you use Claude through the Cowork tab in the Claude desktop app:

1. Open the Cowork sidebar
2. Click Customize
3. Click Browse plugins → Personal → click + → Add marketplace from GitHub
4. Enter https://github.com/panaversity/agentfactory-business-plugins
5. Find IDFA Financial Architect and click Install

The IDFA skill auto-activates in all Cowork sessions when you mention financial models. Both skills — the methodology and the operations — install together as a single plugin.

Plugins in Cowork are saved locally to your machine. For team-wide deployment, your IT team can pre-provision plugins across the organisation through the admin console — see Lesson 10 on governance.

Claude Code (CLI)

If you use Claude Code in the terminal:

/plugin marketplace add panaversity/agentfactory-business-plugins
/plugin install idfa-financial-architect@agentfactory-business

Claude Code reads both SKILL.md files and makes them available in every session. When a conversation mentions financial models, named ranges, or any trigger phrase listed in the skill, Claude activates the IDFA methodology automatically.

Other Agents (GitHub Copilot, VS Code, Codex, Cursor)

The plugin contains a standard SKILL.md file. For agents that do not support the Claude Code plugin format, download skills/financial-architect/SKILL.md from the GitHub repository and place it in the platform's custom instructions path:

Agent	Path
GitHub Copilot	.github/copilot-instructions.md
VS Code	.vscode/copilot-instructions.md
Codex (OpenAI)	Project instructions or system prompt
Cursor	.cursorrules or project-level instructions

The skill content is identical everywhere. You install the plugin once for Claude Code; for other agents, you copy the same SKILL.md from the plugin repo into the platform-specific directory.

Portability

The agentskills.io standard means the methodology is identical everywhere. The plugin packages it for Claude Code and Cowork; the SKILL.md inside the plugin works directly in any other skills-compatible agent. When a new agent announces skills support, installation is a single file copy from the plugin repo.

The Agent Decision Table in Practice

The Agent Decision Table is the operational core of the skill. It tells the agent exactly what to do for every category of financial modelling task. Here is how it works when the skill is active:

Task	What the Agent Does Automatically
Building a new model	Extracts every input from the intent statement, names each with Inp_, writes all calculations in Named Range notation, verifies complex formulas in LaTeX, and attaches Intent Notes
Auditing an existing model	Inspects the model, checks every Calculation layer formula for coordinate references, flags violations, and reports compliance percentage
Retrofitting a legacy model	Inspects the model, identifies all hardcoded values, proposes Named Ranges, rewrites formulas one at a time, and validates that outputs match at each step
What-if analysis	Writes the assumption to the model, lets the spreadsheet engine recalculate, and reads back the results — never calculates internally
Goal-seeking	Iterates writing assumptions and reading results until the target output is reached, then reports the required input value
Explaining a formula	Reads the formula from the model, states the business rule in plain English, and checks for an Intent Note
Checking compliance	Verifies all four guardrails: Named Ranges, LaTeX verification, Intent Notes, and Delegated Calculation workflow

Without the skill installed, you would need to prompt for each of these behaviours explicitly. With the plugin installed, the agent applies the correct workflow based on what you ask it to do.

Trigger Phrases: How the Skill Activates

The description field in the YAML frontmatter lists specific phrases that activate the skill. When a user says any of these, the agent applies the full IDFA methodology:

What You Say	What Happens
"Explain how this model works"	The agent produces a Logic Map in Named Range notation
"This model is a black box"	The agent offers a full audit — inspecting all formulas, inputs, and dependencies
"I inherited this model"	The agent proposes an IDFA compliance audit and retrofitting sequence
"What if revenue grows at 15%?"	The agent writes the assumption to the model, recalculates, and reads back the result
"Check this formula"	The agent verifies Named Range compliance, LaTeX correctness, and Intent Note presence

The trigger phrases are not rigid pattern matches. They are examples that teach the agent the category of request. "This spreadsheet makes no sense" will activate the skill just as "this model is a black box" does — because the agent understands the intent behind the phrase.

The Five Common Mistakes

These mistakes appear in the SKILL.md as a safeguard. When the skill is active, the agent checks its own output against these rules:

1. Never mix layers. A hardcoded 0.60 in a Calculation formula violates layer isolation. The skill teaches the agent to move it to Assumptions as Inp_COGS_Pct_Y1.

2. Never calculate internally. If you ask "what is Year 3 Gross Profit?", the agent must delegate to the spreadsheet engine — never calculate internally. The model's deterministic output and the agent's estimate can differ — and in finance, only the model result is audit-valid.

3. Never skip LaTeX for WACC, IRR, NPV, or DCF. These four formulas are where errors are most common and most consequential. The skill requires LaTeX verification before any of them are committed to the model.

4. Never retrofit by deleting and rebuilding. The skill enforces one-formula-at-a-time conversion with output validation at each step.

5. Never name a range with spaces. Inp_Rev_Y1, not Inp Rev Y1. Excel rejects spaces in formula references; the skill prevents the agent from creating them.

Exercise: Three Ways to Activate the IDFA Skill

You installed the IDFA plugin in the Chapter 18 prerequisites. In Lessons 3-8, the skill auto-activated when you mentioned financial models — you did not need to reference it by name. Now you will learn two more activation methods and understand when each is appropriate.

Step 1 — Prompt: Auto-Activation (What You Have Been Doing)

Open Cowork with a fresh session. Type a financial modelling prompt that mentions IDFA concepts:

Build an IDFA-compliant 5-year SaaS revenue model. Inputs: $5M ARR
growing 30% YoY, 70% gross margins improving to 80% by Year 5.
Use Named Range notation for all formulas. Project revenue, COGS,
and gross profit for all five years.

The skill should auto-activate from trigger phrases like "revenue model," "gross margins," and "Named Range." But auto-activation depends on Cowork recognising those trigger phrases — it is not guaranteed.

Step 2 — Prompt: Explicit Skill Invocation

Now invoke the skill directly by typing /financial-architect in Cowork:

/financial-architect

Audit the SaaS model you just built. Check every formula for Named
Range compliance, verify layer isolation, and confirm Intent Notes
are attached. Report the compliance percentage.

When you type /financial-architect, you are telling Cowork to load the skill explicitly — the full methodology, the Agent Decision Table, the guardrail checks, and the naming conventions all activate at once. This is different from auto-activation: explicit invocation guarantees the skill is loaded, while auto-activation depends on trigger phrases being recognised.

Step 3 — Prompt: Configure the Instructions Pane

Both auto-activation and explicit invocation depend on the agent deciding to load the skill for that particular message. For persistent activation across an entire session, you can configure the Instructions pane to tell Cowork to use IDFA on every interaction.

1. In the Cowork right sidebar, click Instructions (under "outputs") — this opens the Folder Instructions modal
2. In the editor, add this natural-language instruction:

Always apply the IDFA Financial Architect methodology when working
with financial models. Use Named Range notation for all formulas,
verify complex formulas in mathematical notation, attach Intent Notes,
and delegate all calculations to the spreadsheet engine.

3. Save the instructions and return to the chat

This writes to Cowork's equivalent of a CLAUDE.md file — persistent instructions that apply to every message in this session. Unlike auto-activation (which depends on trigger phrases) or explicit invocation (which you must type each time), Instructions pane configuration is always on for the session.

Now test it with a prompt that has no obvious trigger phrases:

Add a customer acquisition cost calculation to the SaaS model.
CAC is $500 in Year 1, decreasing 10% per year. Calculate
payback period using gross profit per customer.

Check the output. Even though the prompt does not mention "financial model" or "Named Ranges," the Instructions pane ensures the IDFA methodology is applied.

Step 4 — Verify: Check the Context Panel

After Cowork processes your prompt, look at the Context section in the right sidebar. When a skill is invoked — whether through auto-activation, explicit invocation, or Instructions — it appears in the Context panel as a loaded skill. This tells you exactly which skills Cowork is using for the current interaction.

In the screenshot above, notice the Context section at the bottom right: it shows financial-architect and idfa-ops as loaded skills, confirming the IDFA plugin is active. The outputs section shows Instructions : CLAUDE.md — the persistent instructions you configured in Step 3.

Verify the output against the four guardrails:

Guardrail	What to Check
Named Range Priority	Are all inputs named with Inp_ prefix? Are all formulas written in Named Range notation with zero coordinates?
LaTeX Verification	If the model includes complex formulas (like payback period), did Cowork verify them in mathematical notation?
Intent Notes	Did Cowork include Intent Note format for generated formulas?
Delegated Calculation	Did Cowork describe the write/read workflow rather than calculating results internally?

Step 5 — Extend: Compare All Three Approaches

Approach	How It Works	When to Use	Reliability
Auto-activation	Trigger phrases in prompt activate the skill	Natural prompts about financial models	High — but depends on phrase recognition
/financial-architect	Explicit skill invocation typed in the chat	Audits, compliance checks, one-off tasks	Certain — skill is explicitly loaded
Instructions pane	Persistent NLP instructions applied to every message	Sessions where every interaction should use IDFA	Always on — no prompt dependency

For everyday model building (Lessons 3-8), auto-activation is sufficient. For formal audits or one-off checks, invoke /financial-architect explicitly. For dedicated IDFA sessions — building a complete model, retrofitting a legacy workbook, or preparing for an audit — configure the Instructions pane so the methodology is active without relying on trigger phrases.

The Business Bottom Line

One plugin install. Every agent. Every model interaction governed by the same four guardrails. For non-Claude agents, the same SKILL.md file from the plugin repo installs directly into the platform's custom instructions path.

Everything you learned in Lessons 1 through 8 is knowledge in your head. The IDFA plugin is knowledge in every agent you use. The investment is a single install command — the return is consistent IDFA compliance across every tool that supports the agentskills.io standard.

This also connects back to Chapter 15, where you learned the plugin architecture for Agent Skills. The IDFA plugin is not a custom configuration — it is a standard-format skill that any skills-compatible agent can read. The same investment that makes Claude apply IDFA also makes Copilot, Codex, Gemini CLI, and future agents apply it. The skill travels with your project, not with any single tool.

Try With AI

Setup

Use these prompts in Cowork or your preferred AI assistant with the IDFA plugin installed (see Chapter 18 prerequisites if you have not installed it yet). For the most reliable results, configure the Instructions pane as described in the exercise above. Start a fresh session so the skill is loaded.

Prompt 1 — Test skill activation with a new intent statement:

I need to build a financial model for a hardware startup.
Year 1 unit sales are 10,000 at $200 per unit. Unit sales grow
25% YoY. Manufacturing cost is $120 per unit in Year 1,
decreasing 5% per year through supply chain optimisation.
Project revenue, COGS, and gross profit for 5 years.

What you are learning: Whether the skill activates automatically from a financial modelling request. Check the output for Inp_ prefixed Named Ranges, Named Range-only formulas, and the delegated calculation workflow. If the agent produces coordinate-based formulas or calculates results internally, the skill is not active — verify the plugin is installed and restart the session.

Prompt 2 — Test trigger phrase activation:

I inherited this model from a colleague who left the company.
It has 300 formulas across 8 tabs. I need to understand what
it does before the board meeting next week.

What you are learning: Whether trigger phrases activate the correct IDFA workflow. The phrase "I inherited this model" should trigger the agent to offer an IDFA compliance audit, inspect the model to map all formulas, and suggest a retrofitting sequence. If the agent responds with generic spreadsheet advice instead of the IDFA methodology, the skill's trigger phrases are not being read.

Prompt 3 — Build a complete model with the skill active:

Build a 3-year consulting firm profitability model.
Inputs: 50 consultants at $150K average salary, growing to
65 consultants by Year 3. Average billable rate $250/hour.
Utilisation rate starts at 70% and improves to 78% by Year 3.
Assume 1,800 available hours per consultant per year.
Project revenue, personnel costs, gross profit, and
gross margin percentage for all three years.

What you are learning: Whether the skill governs a complete model-building workflow end to end. The output should include: every input extracted and named with Inp_, all calculation formulas in Named Range notation, LaTeX verification for any multi-step formulas, Intent Notes for each generated formula, and the delegated calculation workflow. This is the full test — if all four guardrails appear without prompting, the IDFA skill is fully operational.

---

Next: Lesson 10: Enterprise Governance — where you design the organisational controls that the plugin cannot enforce: accountability, tracking, approval workflows, and compliance metrics.