1 Actionable AI Workflow

A proper 3-statement operating model is one of the most powerful tools in fundamental investing.

It’s also one of the most painful to build. Income statement. Balance sheet. Cash flow. Linked assumptions. Forward projections.

Done right, it’s what separates a real investment view from a guess. Done manually, it takes a skilled analyst days.

I built one in about an hour using AI. Here’s exactly how.

The workflow: Manus to assemble the data room + pull historical assumptions, then Claude for Excel to construct the full model, then Claude Code to double check the work.

The stock I modeled: ComfortDelGro (SGX: C52).

I picked this coming off a month in Asia working with hedge funds and investment banks. One theme kept surfacing at the CIO level that I wanted to dig deeper on: Singapore’s Equity Development Programme (EQDP) - a government-backed S$6.5 billion deployment into fund managers targeting Singaporean equities.

(BONUS! I put together a full 32-page research dossier on investing off EQDP .... Download it here or below EQDP_Research_Dossier.pdf)

Bonus - Download EQDP Dossier

Why the EQDP is interesting is straightforward. Capital inflow → improved liquidity → rising volumes → institutions can deploy at size → valuation re-rating on small and mid-caps. For individual investors who aren’t capacity-constrained, this is a structural edge. You can front-run institutional inflow before the liquidity is fully there.

But conviction on individual names requires a model. Not a rough napkin math exercise… a real one.

And for readers outside the US who keep asking whether my AI techniques work on stocks ex-USA: ComfortDelGro trades on the SGX, one of the smaller exchanges in the world. If it works here, it works anywhere.

The Prompt(s)

We are building this 3 statement operating model with 3 prompts.

Preparing Data Room

The first prompt fetches the data room of all the company documents in order to build this model (earnings releases, material statements, official financials, company presentations etc.). We will be using Manus for this task to prep a .zip file.

Prompt 1:Data_Room_Prompt.pdf

We fetched a ~400 document data room to work with! This increases data fidelity.

Fetching Assumptions for the Model

In the same chat, now we want to fetch the assumptions so we can properly RAG this into Claude for Excel. This protects context windows.

Prompt 2: Manus_Fetch_Historicals_Assumptions.pdf

Result: ComfortDelGro_Assumptions.pdf

Building the Model

Now in Claude for Excel, we use this detailed prompt alongside the above assumptions doc we prepared.

Prompt 3: PROMPT_FOR_CLAUDE_EXCEL.pdf

The Result

Full output here:ComfortDelGro_Op_Model.xlsx

The output is a fully linked 16-sheet, 3-statement model for ComfortDelGro -- income statement, balance sheet, cash flow, valuation, and sensitivity tables. FY2020 actuals through FY2029E projections. Revenue built bottom-up across 5 segments.

But getting the numbers right required one more step.

Claude Code + the data room as a fact-checker.

We ran Claude Code against the full ~400-document data room saved locally to cross-check every historical figure against primary filings. One catch worth knowing: Claude Code works in Python and hardcodes outputs, so you can't directly pipe results back into Claude for Excel. You need deliberate logic to transfer verified figures between the two environments.

Worth it though. In ComfortDelGro's case, the cross-check caught historical financial restatements the original assumptions had missed. Those errors would have silently corrupted every downstream projection.

Is this model perfect yet? No.

We only have starting assumptions in the model. The real work -- and where you actually get paid as an investor -- is interrogating further the business drivers:

Can Public Transport margins sustain their recovery through the contract renewal cycle?
How much structural decline is left in Taxi/PHV vs. Grab?
What's the true run-rate contribution from A2B Australia?

AI gets you to those questions faster. The model infrastructure is no longer the bottleneck. Your thinking is.

The important thing is we now have more time to do "thinking work" by leveraging AI for the "monkey work"

Claude Code for Finance LIVE Webinar

What you saw in this newsletter was powered by Claude Code.

It's the most powerful AI tool I've used for finance.

I'm putting together a live webinar course designed specifically for investors and buyside professionals. Here's what we'll cover:

Setting up Claude Code from scratch (zero terminal experience required)
Navigating the interface for finance workflows
Asset research + data room assembly
Investment memo writing
Excel modeling and valuation
Forensic accounting and filing analysis
Live Q&A on your specific use cases
...and more

This will be a capped live webinar course. Spots will fill fast once we open registration.

Reply to this email or click the waitlist link below to get notified as soon as it drops. Hope to see you then.

Join the Waitlist

Get one high impact AI prompt every week