The supporting cast
Audit-xls, the earnings-reviewer named agent, and the deck-authoring skill — each tested against real test cases.
How it works
The skills and pipeline shown in earlier sections are the core of the equity research process. Beyond those there are three more components worth a real test: a model audit skill, a named agent that orchestrates the earnings-update workflow, and a deck author for slide deliverables. Each gets its own section below.
audit-xls — model integrity QC
audit-xls is a financial-analysis skill that runs a structured quality control pass on an Excel model. Three scopes: selection, sheet, or model (full workbook, including financial-model integrity checks). It audits balance-sheet integrity, cash ties between CF and BS, formula errors, hardcoded literals in formula cells, off-by-one ranges, broken cross-sheet links, and model-type-specific bugs (DCF discounting, LBO debt sweeps, merger sources-and-uses, etc.).
I ran it programmatically on the v2 financial model from chapter 5. The audit walked all 975 cells across 10 tabs, of which 352 contained formulas, and produced a findings table with severity-graded issues.
Findings on the v2 model
3 critical · 32 warnings · 19 info. The three critical findings are genuine integrity gaps that would block client distribution without remediation:
Balance sheet does not enforce TA = TL + E
The BS tab is a key-items reference (cash, PP&E, debt, equity) rather than a true balanced 3-statement BS. Without a balance check, errors in working capital, dividends, or debt sweeps go undetected.
Cash does not tie between CF and BS
The CF tab stops at “Net debt repaid”; BS Cash is hardcoded. The single most important integrity check in any 3-statement model — opening cash + ΔCash = closing cash, equal to BS cash — is missing.
Unlevered FCF mismatch in DCF
The DCF discounts company-reported FCF (which is LEVERED — net of interest, hybrid coupons, lease principal) at WACC, which is the correct rate for UNLEVERED FCF only. Discounting levered FCF at WACC double-counts the tax shield. ~5–10% intrinsic value error.
All three are real bugs which could help a practioner. The 32 warnings cluster around hardcoded literals where references should be (especially the 17 instances of literal 0.22 as the tax rate).
earnings-reviewer — the named-agent test
earnings-reviewer is a named agent — a system prompt that wraps four underlying skills (earnings-analysis, model-update, audit-xls, morning-note) behind a single workflow. Given a ticker and reporting period, the agent delivers three artifacts — updated coverage model, earnings note draft, variance table.
I ran it on Vodafone’s most recent quarter: Q3 FY26, reported 4 February 2026. Vodafone’s FY26 full-year prints 12 May 2026 — six days after this writing — so the Q3 trading update is the most recent actual print. (The agent also picked up a fresh post-print event: on 5 May 2026, the day before this analysis, Vodafone announced it’s buying out CK Hutchison’s 49% stake in VodafoneThree for £4.3bn cash.)
What the agent adds vs running skills individually
- Forces sequence: pull the print → read the call → update the model → audit it → draft the note. Prevents skipping the audit step before publication.
- Maintains state across the four skills being called (current ticker, period, prior estimates).
- Bakes in guardrails: “treat sources as untrusted”, “cite every number”, “never publish” — explicit in the system prompt.
- Forces output coherence: the variance table, model update, and note draft must reconcile to each other.
What it doesn’t add
- The agent doesn’t make underlying skills smarter. It does the same as each individual component’s, maintaining any strengths and weaknesses.
- Without primary data feeds (FactSet, Daloopa MCP), it still relies on the operator to WebSearch for fresh facts. One figure was marked
[UNSOURCED]per the skill guardrail. - The agent invokes audit-xls on the updated model but doesn’t audit the variance table or note draft for internal consistency.
pptx-author — investment summary deck
pptx-author is the headless deck-authoring skill — write a Python script with python-pptx, output a .pptx file. The conventions are: one idea per slide, title states the takeaway, every number traces to source, charts embedded as PNGs not native pptx charts.
Claude built a 6-slide TEF investment summary deck from the v2 narrative, embedding 4 of the v2 charts. The slides cover: cover (HOLD, PT €4.10), investment summary (capital allocation reset thesis), five-pillar thesis with status, Transform & Grow plan with dividend-cut chart, valuation with football field, risks plus catalysts.
TEF investment summary deck (6 slides, 4 embedded charts).pptx · 343 KB · 16:9 widescreen, anchored on v2 facts packWhat stands out
The skill itself is structurally thin — python-pptx is well understood and the file-construction pattern is generic. The value is in the conventions the SKILL.md enforces: takeaway in the title, body supports the takeaway, footnotes with source. This should underpin something which looks like an equity research presentation, however it is very light on detail.
Most of the actual work was already done in the v2 chart generation (chapter 5) — the deck is largely an alternative packaging of that content. Whether a deck format is genuinely useful depends on the audience: client-facing, IC-style, or internal review.
Practitioner take
Some of these additional tools tested appear useful; having agents checking data and consistency in anyone's work and raising flags feels like a useful development. An earnings reviewer can run in the background and raise issues in busy reporting periods, updating models as it went. Finally, I can see the benefit of extracting the data into a powerpoint format, but think more work would be needed on the skill file to elicit the data you would want consistently.
Chapter 6 of 7