Extracting Finance Domain Knowledge

In Lesson 4, you installed the knowledge-work-plugins/finance plugin and saw what a generic finance agent can do: standard close workflows, industry-standard variance analysis, and templated management reporting. Now you will see what it cannot do — and why the gap matters.

The generic plugin applies the same materiality thresholds to every company. It uses textbook variance analysis that treats a five percent deviation in revenue the same way it treats a five percent deviation in a dormant travel expense account. It produces management bridges in a standard format that may not match the format your CFO presents to the board. The plugin is useful. It is not yours.

The gap between "useful" and "production-ready" is filled by the kind of knowledge that lives in the CFO's head: which variances actually matter at this firm, what seasonal patterns distort standard analysis, which accounts deserve scrutiny regardless of their dollar magnitude, and what the board expects to see in the variance bridge. This lesson applies the Knowledge Extraction Method from Chapter 16 to surface that knowledge, and walks through the translation from extraction outputs to a first-draft SKILL.md that encodes firm-specific close judgment.

Why Generic Plugins Fall Short

A generic finance plugin handles the mechanics of the monthly close competently. It can calculate variances, apply standard materiality thresholds, and produce formatted reports. What it cannot do is apply the judgment that an experienced CFO brings to the same data.

Consider variance analysis. The generic plugin flags any line item that deviates from budget by more than a fixed percentage — say, ten percent or $100,000, whichever is greater. An experienced CFO at a mid-market manufacturing company applies different logic entirely:

Variance	Generic Plugin Response	CFO's Actual Response
Revenue down $50K against a $20M budget	Flags for review (exceeds threshold)	Ignores — within normal monthly noise for this business
Travel expense up $5K in a dormant cost centre	Ignores — below threshold	Investigates immediately — dormant accounts with activity are a potential fraud signal
COGS up 8% in Q4	Flags as unfavourable	Expected — seasonal raw material pricing; the CFO adjusts the baseline before comparing
Deferred revenue down 15%	Flags as unfavourable	Triggers a call to the VP of Sales — at this firm, deferred revenue drops precede revenue misses by one quarter

The CFO's responses are not random departures from the standard. They are the product of years of experience with this specific business, its seasonal patterns, its risk profile, and its historical failure modes. This is the tacit knowledge that Chapter 16's extraction methodology is designed to surface.

The Five Questions Applied to Finance

Chapter 16 Lesson 2 introduced the Five Questions that structure a Method A interview. Each question targets a different layer of tacit knowledge. Here is what each one surfaces when applied to a CFO's monthly close workflow.

Question 1: What do you do that no manual or policy document captures?

This question surfaces the undocumented heuristics that govern the CFO's real decision-making. In finance, the answers tend to cluster around three areas: materiality judgment, account-level attention patterns, and timing logic.

Example extraction material: The CFO explains that a $50K variance in revenue is noise at their firm — the business has enough monthly volatility that anything under $75K does not warrant investigation. But a $5K variance in a dormant account gets immediate attention, because dormant account activity outside of year-end reclassifications has historically been an indicator of unauthorised transactions or misclassifications that compound over quarters if uncaught.

This is a candidate SKILL.md Principle. The generic plugin treats both variances identically against its fixed threshold. The CFO's logic is specific, testable, and grounded in the firm's actual risk profile.

Question 2: When do you override the standard process?

This question surfaces the conditions under which the CFO's judgment replaces the documented procedure. In finance, overrides cluster around non-standard periods — acquisitions, restructurings, system migrations, and seasonal distortions.

Example extraction material: The CFO explains that for the first three closes after an acquisition, they waive normal materiality thresholds for the acquired entity's accounts. Integration adjustments — purchase price allocation entries, system migration reclassifications, one-time transition costs — distort the standard baselines so severely that applying normal variance analysis produces hundreds of false positives. Instead, they run the acquired entity on a separate review cadence with wider thresholds until the baselines stabilise, typically after the third full quarter.

Question 3: What signals tell you something needs attention before the numbers confirm it?

This question surfaces leading indicators — the early warning patterns that experienced finance professionals detect before they appear in the standard metrics.

Example extraction material: The CFO watches accounts receivable aging shifts as a leading indicator of customer credit stress. Even when the days sales outstanding number looks acceptable in aggregate, a shift in the aging distribution — specifically, the 60-90 day bucket growing as a percentage of total receivables while the 0-30 day bucket shrinks — signals that collection patterns are deteriorating. The DSO metric catches this two to three months later. The aging distribution catches it now.

Question 4: What mistakes do new team members make that experienced ones avoid?

This question surfaces the experience gap — the errors that training does not prevent because they require context that only comes from working in the specific environment.

Example extraction material: The CFO describes a recurring pattern: new controllers post intercompany elimination entries before confirming that the subsidiary has completed and locked its close. The subsidiary's numbers shift, the elimination entries become wrong, and the consolidation has to be reversed and re-run. Experienced controllers always confirm subsidiary close status before touching eliminations — but this sequence is not documented in any procedure manual because it seems obvious to anyone who has been burned by it once.

Question 5: What would you tell your replacement on their first day?

This question surfaces the distilled wisdom — the load-bearing instructions that the expert considers most critical.

Example extraction material: The CFO says: "Never trust the revenue number until you have confirmed the deferred revenue schedule is clean. At this company, revenue recognition depends on milestone completion in our professional services contracts, and the milestones are confirmed by project managers who are not accountants. Every month, at least two milestones are marked complete when the deliverable is still in client review. If you take the revenue number at face value before checking the deferred revenue schedule against the milestone tracker, you will report revenue that has to be reversed next month."

From Interview to SKILL.md Draft

The extraction material above — gathered from a single sixty-minute interview — provides enough raw material to draft the three sections of a monthly-close-judgment SKILL.md.

The Persona

Using the three Persona writing questions from Chapter 16 Lesson 6:

What is this agent's professional level and authority? A senior controller with ten years of experience at this specific firm, reporting directly to the CFO. The agent operates with the authority to approve standard close entries and flag non-standard items for CFO review. It does not have authority to approve adjustments above $250K or to override established accounting policies.

What does this agent value in its outputs? Precision over speed. An incomplete close that identifies the three items requiring CFO attention is more valuable than a complete close that buries an unresolved item in the notes. The agent prioritises catching the item that will cause a restatement over meeting the close deadline by two hours.

What does this agent do when it does not know? When the agent encounters a variance it cannot classify using the firm's established patterns — a new account type, an unfamiliar transaction pattern, or a variance that does not match any seasonal or operational precedent — it flags the item explicitly, describes what it observed, states that it has no precedent for this pattern, and routes to the CFO for judgment. It does not approximate an answer.

The Questions Section

In scope:

• Monthly close variance analysis using firm-specific materiality thresholds
• Subsidiary close status verification and intercompany elimination sequencing
• Deferred revenue schedule reconciliation against the milestone tracker
• Seasonal baseline adjustments for Q4 raw material pricing and Q1 contract renewals
• Variance bridge preparation in the CFO's board presentation format

Out of scope:

• Accounting policy changes or new standard interpretations (route to external auditors)
• Close entries for the first three months following an acquisition (separate review cadence)
• Variance analysis for entities in active restructuring (CFO handles directly)
• Any external financial reporting or communications (requires CFO review and sign-off)

The Principles (Selection)

These Principles are drafted directly from the extraction material. Each one passes the testability criterion — you can construct a scenario and confirm the agent follows the instruction.

1. Dormant account scrutiny: When any account classified as dormant shows activity above $1,000 outside of year-end reclassification periods, flag for immediate investigation regardless of the firm-wide materiality threshold. Dormant account activity is a potential indicator of unauthorised transactions.

2. Deferred revenue verification: Before accepting the revenue figure for close, reconcile the deferred revenue schedule against the professional services milestone tracker. When any milestone is marked complete but the client deliverable status shows "in review" or "pending approval," treat the associated revenue as unconfirmed and flag for the CFO.

3. Subsidiary close sequencing: Do not post intercompany elimination entries until the subsidiary has confirmed its close is complete and locked. Confirm lock status directly — do not infer from the subsidiary's reported completion date.

4. Post-acquisition threshold override: For the first three complete quarters following an acquisition, apply a separate materiality threshold to the acquired entity's accounts. The threshold is set by the CFO at integration kick-off. Do not apply the firm's standard thresholds during this period.

5. Aging distribution as leading indicator: When the 60-90 day bucket of accounts receivable increases by more than 5 percentage points as a share of total receivables over two consecutive months, flag as a potential credit stress signal even if the aggregate DSO metric remains within acceptable range.

This draft is a first draft — it encodes the extraction material faithfully but has not been tested against the range of scenarios the agent will encounter. The Validation Loop from Chapter 16 Lessons 7 and 8 is the next step.

Extracting from Close Documents

Method B — document extraction — complements the interview by surfacing knowledge that lives in institutional documents rather than in the expert's memory. For a finance close workflow, three document categories contain extractable knowledge.

The close checklist. Apply the three-pass framework from Chapter 16 Lesson 4. Pass One extracts the explicit rules — the sequence of close activities, the sign-off requirements, the deadline structure. Pass Two identifies contradictions between the documented procedure and actual practice — the CFO's interview may reveal that certain checklist items are always done out of order or that some sign-offs are treated as formalities. Pass Three identifies gaps — close scenarios that the checklist does not address, such as how to handle a close when the subsidiary's ERP system is being migrated.

The accounting policies manual. This document contains the firm's interpretations of accounting standards — revenue recognition policy, capitalisation thresholds, lease classification criteria. The explicit rules from Pass One translate into Principles. The contradictions from Pass Two — where practice diverges from documented policy — are some of the most valuable extraction material because they reveal where professional judgment fills the gap between policy and reality.

The prior period adjustments log. This document records every adjustment that was made after a close was initially completed. Each adjustment represents a case where the standard process produced an incorrect result. The pattern of adjustments reveals the recurring failure modes that the SKILL.md's Principles should prevent.

Bridge to Enterprise Extensions

The SKILL.md you drafted in this lesson targets one workflow: the monthly close. Lessons 10 and 11 introduce eleven enterprise extension areas — credit risk, regulatory reporting, treasury management, FP&A and budget ownership, and seven more — each of which follows the same extraction pattern.

Every extension begins the same way: identify the firm-specific knowledge that a generic plugin cannot encode, extract it using the Five Questions and document analysis, and translate it into a SKILL.md that makes the agent yours. The close workflow is the first application. The extensions are the roadmap.

Try With AI

Use these prompts in Anthropic Cowork or your preferred AI assistant to practise the finance extraction skills from this lesson.

Prompt 1: Finance-Specific Five Questions

I need to prepare for a knowledge extraction interview with a
[FINANCE ROLE: e.g., CFO, VP of Finance, Head of FP&A, Treasury
Manager] at a [COMPANY TYPE: e.g., mid-market SaaS company,
manufacturing firm, professional services firm].

For each of the Five Questions from the Knowledge Extraction Method,
generate:
1. The finance-specific version of the question adapted to this role
2. Two example answers the expert might give — one generic (not useful
   for SKILL.md) and one specific (directly translatable into a
   Principle)
3. A follow-up question that would move a generic answer toward a
   specific one

Focus on the workflows this role owns and the judgment calls that
differ from textbook practice.

What you're learning: The Five Questions are a framework, not a script. Adapting them to a specific finance role requires understanding what tacit knowledge that role holds. A CFO's undocumented heuristics differ from a Treasury Manager's — the same question structure surfaces different material. Practising the adaptation builds the flexibility to extract from any finance role.

Prompt 2: Generic-vs-Specific Variance Analysis

Here are five statements a CFO might make during a knowledge extraction
interview about their monthly close process. For each statement:

1. Classify it as generic or specific
2. If generic, write the follow-up question that would surface the
   specific version
3. If specific, draft the SKILL.md Principle it would produce

Statements:
- "We always investigate material variances"
- "When seasonal inventory builds in Q3 exceed the forecast by more
  than 12%, I call the supply chain VP before adjusting the reserve
  because the excess is usually pre-positioned for Q4 holiday orders"
- "Experience teaches you which numbers to trust"
- "I never close the books until AP confirms all invoices above $25K
  are posted, because we had a quarter where a missing invoice caused
  a material understatement of COGS"
- "We follow GAAP carefully"

What you're learning: The generic-to-specific transformation is the core extraction skill applied to finance. Statements like "we follow GAAP carefully" contain no actionable knowledge for a SKILL.md. Statements like the seasonal inventory build contain a specific condition, a specific threshold, a specific action, and the reasoning behind it — everything needed for a testable Principle. Training this distinction with finance-specific material builds the pattern recognition you need during a real extraction interview.

Prompt 3: SKILL.md Draft from Interview Notes

I conducted a knowledge extraction interview with a [FINANCE ROLE]
at [COMPANY TYPE]. Here are my interview notes:

[Paste your actual interview notes, or use the following practice
scenario: A VP of FP&A at a $200M SaaS company who manages the annual
budget process and monthly forecast updates. Key extraction material:
they re-forecast quarterly rather than monthly for headcount because
hiring plans change too slowly to justify monthly updates, they use
bookings pipeline coverage ratio as a leading indicator for revenue
forecast accuracy, and they never present a budget variance without
the bridge showing whether the variance is timing or permanent.]

Help me draft the three sections of a SKILL.md:
1. Persona — using the three Persona writing questions
2. Questions — with both in-scope and out-of-scope boundaries
3. Principles — at least five, each passing the testability criterion

After drafting, assess each Principle: Is it specific enough to test
against a scenario? If not, identify what additional interview
question would surface the missing specificity.

What you're learning: The full pipeline — from interview notes to SKILL.md draft — is the synthesis skill this lesson builds. The assessment step at the end trains you to evaluate your own extraction quality: Principles that feel reasonable but fail the testability criterion reveal where the interview did not go deep enough. This is the feedback loop that improves both your extraction technique and your SKILL.md authorship with each iteration.

Flashcards Study Aid

---

Continue to Lesson 9: Enterprise Extensions — Risk and Compliance →