Corporate credit risk analysis is one of the most document-heavy workflows in banking. A single mid-market borrower can generate a 60-page annual report, two years of management accounts, a covenant compliance certificate, an auditor’s letter, and a dozen email threads with the relationship manager. Before a credit committee even meets, an analyst must reconcile all of it into a coherent view of repayment capacity, leverage trajectory, and downside exposure. The bottleneck is rarely analytical skill. It is the mechanical work of reading, extracting, cross-checking, and formatting — work that consumes days and introduces transcription errors that no one catches until the committee asks a hard question.
The friction compounds at portfolio level. A CFO or controller preparing a quarterly credit review does not need one borrower analyzed well; they need forty borrowers analyzed consistently, with the same ratios, the same definitions, and the same risk flags applied to each. Consistency is where manual processes break down first. Two analysts reading the same balance sheet will classify a convertible note differently, adjust EBITDA differently, and reach different conclusions about the same covenant headroom.
ChatGPT closes most of that gap when it is given structure. It will not replace judgment — it should not be asked to. But it will read every page, extract the figures you specify, apply your definitions without drift, and produce a first draft that a credit professional can review in an hour instead of building in a day. The two prompt templates below are designed for exactly that: a borrower-level risk memo and a portfolio-level screening pass. Both follow the same discipline — load the source files, define the output, force the model to ask questions before it writes anything.
Why structure beats cleverness in credit prompts
Credit risk prompts fail for predictable reasons. The model invents ratios that were never calculated, mixes fiscal years, or produces a polished narrative that quietly omits the one disclosure that matters. The fix is not a more sophisticated prompt. It is a prompt that forces the model to declare its plan, state its assumptions, and confirm which documents it has actually read before it produces a single number. Every figure in the output should be traceable to a page in a source file.
The templates that follow build that discipline in. Note the instruction to ask clarifying questions first — in credit work this is not a courtesy, it is a control. If the model does not know whether your leverage definition includes operating leases, it should say so rather than pick one silently.
First, read these files completely before responding:
[borrower_annual_report_2025.md] — audited financial statements, notes, and management discussion
[borrower_management_accounts_q2_2026.md] — unaudited interim P&L, balance sheet, and cash flow
[covenant_certificate_q2_2026.md] — facility agreement covenant definitions and compliance calculations
[internal_credit_policy.md] — our ratio definitions, risk grading scale, and approval thresholds
Here is a reference for what I want to achieve:
[Upload a previously approved credit memo as markdown]
Here’s what makes this reference work:
It states the facility structure before the analysis. It presents every ratio with the formula, the inputs, and the source document. It separates factual findings from analyst judgment under distinct headings. It flags covenant headroom in absolute currency terms, not just as a ratio. It ends with a recommendation, a risk grade, and the specific conditions or monitoring triggers attached to that grade.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: Credit risk memo, 1,200 to 1,800 words, with a ratio table
Recipient’s reaction: The credit committee should be able to approve, decline, or approve-with-conditions without requesting additional analysis
Does NOT sound like: A sell-side equity research note, or a generic summary that restates the financials without interpreting them
Success means: Every ratio in the memo traces to a named source document and page, and the risk grade recommendation matches our internal policy thresholds
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Run that prompt against a borrower you already know well the first time. You will immediately see where the model’s ratio definitions diverge from your policy, and you can correct the context file rather than the output. After two or three calibration runs, the first draft becomes genuinely usable.
Scaling the same discipline to a portfolio
Single-borrower memos are the visible work. The hidden work is portfolio surveillance — spotting which of your forty exposures has deteriorated since last quarter. This is a different task with a different output: shorter, comparative, and ranked. The prompt below produces a screening table rather than a narrative, because at portfolio level you need triage, not prose.
First, read these files completely before responding:
[portfolio_exposure_list.md] — borrower names, facility types, limits, outstandings, and internal risk grades
[borrower_financials_batch_q2_2026.md] — interim financials for each borrower in a consistent format
[watchlist_criteria.md] — our defined early warning indicators and their thresholds
[prior_quarter_screening.md] — last quarter’s screening output for comparison
Here is a reference for what I want to achieve:
[Upload last quarter’s screening output as markdown]
Here’s what makes this reference work:
It ranks borrowers by severity rather than alphabetically. Each row shows the trigger, the threshold, the actual value, and the quarter-on-quarter direction. It distinguishes a genuine deterioration from a seasonal pattern. It names the single most important question the relationship manager should answer for each flagged borrower.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: Screening table plus a 400-word summary of the three most urgent exposures
Recipient’s reaction: The head of credit should be able to assign review owners directly from the table without further triage
Does NOT sound like: An alarmist list that flags every minor variance, or a data dump with no prioritization
Success means: Every borrower above threshold appears in the table with a traceable trigger, and no borrower below threshold is included
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Two practical notes before you deploy either template. First, keep your internal credit policy in a versioned markdown file and update it whenever definitions change — the model’s consistency is only as good as the file it reads. Second, never let the model’s output reach a committee without a named analyst signing off on the figures. The value here is speed and consistency, not autonomy. Use the saved time to interrogate the assumptions rather than to skip them.
A reasonable next step is to run the portfolio screening prompt on a historical quarter where you already know the outcome, and compare the model’s flags against what actually happened. That single backtest will tell you more about where to set your thresholds than any amount of prompt refinement.
Published on 27 September 2026 on growwithgpt.com