Most collections of finance prompts have the same problem: every prompt makes you re-type your business. Revenue here, currency there, country, headcount, structure. By the third one you stop bothering, and the answers get worse because Claude is guessing at everything you didn't repeat.
So this page is built the other way round. You fill in one context block, once, at the top of a chat. Every prompt after it is four or five lines, because Claude already knows who you are. The whole thing takes about ten minutes the first time and about ninety seconds every time after.
I'm Gašper. I built Zenith because European businesses have their numbers sitting in a bank account nobody queries until year-end. These prompts work with any figures you paste in — you don't need Zenith. If you do connect it, Claude reads your live transactions through MCP and most of the pasting disappears. I'll show both versions.
Step 1: The context block
Paste this at the top of a new Claude chat and fill it in. Everything below inherits it.
Business context for this conversation. - Business: [what we do, in one line] - Country of incorporation: [X]. We sell into: [countries] - Structure: [limited company / sole trader / group] - Revenue last 12 months: [amount]. People: [X] - Reporting currency: [X]. Other currencies we hold or invoice in: [X] - VAT registered in: [countries] - Accounting system: [X]. Banks: [X] - Month-end close happens: [when] - My role: [owner / finance lead] - The financial thing I'm most worried about right now: [one sentence] How I want you to work: 1. Use my actual figures. If a number you need is missing, ask for it instead of assuming one. Never fill a gap with a plausible placeholder. 2. Show your working on every calculation, and list your assumptions at the end of each answer. 3. Convert everything to my reporting currency. State the rate and the date you used. 4. If an answer depends on national rules, name the country whose rules you applied and tell me to verify it. 5. When I ask for a file, build a real .xlsx or PDF. Don't paste a table into the chat and call it a report. 6. Be direct. If the numbers are bad, say the numbers are bad. Confirm you've got this, then wait for my first question.
Rule 1 is the one that matters. Left alone, Claude will quietly invent a figure to keep an analysis moving, and a forecast built on one invented number looks exactly like a forecast built on real ones. Telling it to ask instead is the difference between a report you can act on and a very confident piece of fiction.
With Zenith connected: rules 1 and 3 mostly take care of themselves, because Claude is reading actual transactions in actual currencies rather than working from what you remembered to paste.
Step 2: How your numbers get there
| Paste it manually | Connect it live via MCP | |
|---|---|---|
| Setup | None | ~5 minutes, once |
| Before each question | Export, clean, paste | Just ask |
| Data freshness | Your last export | Current bank feed |
| Categorization | Whatever the export has | Consistent month to month |
| Survives as a habit? | Rarely | Yes |
Pasting works. It just doesn't last — almost nobody is still exporting CSVs in month three, which is why most owners run this kind of analysis once and never again. Zenith's MCP connection is read-only and takes about five minutes to set up.
Prompt 1: Cash position and runway
How much do I actually have, and how long does it last?
Run a cash position review. Balances: [account: amount] for each account Credit facility available: [amount] Average monthly burn: [amount] Committed payments in the next 60 days: [amount, what, when] Give me total available cash, runway in months at current burn, and the same again after the committed payments. Tell me the calendar date I'd need to act by, not just a month count. Flag any balance sitting idle. Build it as a one-page PDF summary, dated today.
Live version: "What's my cash position across all accounts including facilities, and my runway at current burn? One-page PDF."
Watch for: the date is the output that changes behaviour. "Seven months" gets nodded at; "decide by 14 March" goes in the calendar.
Prompt 2: Spend and subscription leaks
What's quietly renewing that nobody approved?
Forensic spend audit on these transactions: [paste last 6 months] Find: every recurring charge grouped by vendor with monthly and annual cost; any vendor up more than 20% versus the prior period, and whether that's price or volume; overlapping tools doing the same job; charges that should have stopped when a project ended or someone left. Excel workbook: one row per issue, columns for vendor, monthly cost, annual cost, amount at stake if I act, and a one-line action. Sort by annual amount at stake. Add a total row.
Live version: "What am I spending on SaaS, and has anything crept up in the last three months? Anything worth cancelling goes in a spreadsheet."
Watch for: work down the list until the annual saving drops below what an hour of your time is worth. The first three rows are usually most of the money.
Prompt 3: 13-week cash flow forecast with stress tests
What does cash look like next quarter if things go badly?
The most useful spreadsheet a growing business owns, and almost nobody under €5m revenue has one.
Build a 13-week rolling cash flow forecast. Opening balance: [amount] Weekly cash in and out, last 26 weeks: [paste] Largest 5 expected receipts and 5 expected payments, with weeks: [...] Recurring weekly outgoings: [amount] Tax payments due, with dates: [...] Facility headroom: [amount] Week by week: receipts, payments, net, closing balance, weeks of cover. Flag every week that drops below 4 weeks of operating cost. Then stress test: (a) revenue down 20% from week 3, (b) my three biggest customers each pay 60 days late, (c) both. For each, give the lowest cash point, the week it lands, and whether the facility covers it. Finish with three things to do this week if (c) looks plausible. Excel file: one tab per scenario, plus a summary tab with all four balance lines on one chart.
Live version: the 26 weeks of history is the part people give up on. It's already in your bank feed, so Claude pulls it instead of you rebuilding it from statements.
Watch for: if scenario (c) goes below zero in any week, that's not a forecasting problem, it's a working capital problem. Go to Prompt 5 this week.
Prompt 4: Concentration risk
Am I dangerously dependent on one customer or one supplier?
Concentration risk scan. Revenue by customer, last 12 months: [paste] Costs by vendor, last 12 months: [paste] Any top-3 supplier that's single-source: [...] Percentage of revenue from my top 1, 3 and 5 customers. Same for costs and vendors. Tell me whether that's beyond healthy for my size, using a stated threshold and saying where it comes from. Then model it: if my largest customer leaves, what happens to annual profit and how many months of runway does that cost me? Flag any single-source supplier whose failure would stop delivery, not just raise costs. One-page PDF with a chart for revenue concentration, one for costs, and the "largest customer leaves" number in a callout.
Live version: who pays you and who you pay is already organised in the feed, which is most of the assembly work this needs.
Watch for: the percentage doesn't move anyone. The profit number does. "Client A is 34% of revenue" versus "losing Client A costs €81k and four months of runway" are the same fact with very different consequences.
Prompt 5: Who pays you late, and what it costs
Which customers are financing themselves with my cash?
This is the prompt that finds money without selling anything. Every business has two or three customers who have quietly decided your terms are a suggestion, and almost nobody has measured what that habit costs.
Payment behaviour audit. Money received, last 12 months — date, amount, counterparty: [paste] My standard payment terms offered: [X days] Money paid out to suppliers, same period: [paste] What I'd pay to borrow: [X%] Invoice issue dates, if I have them to hand: [paste — optional] For each recurring customer: average gap between when payment was due and when the money actually landed, how variable that gap is, and whether it's getting better or worse across the year. Rank customers by days late and by euros tied up. Tell me what one day of delay costs me across all customers at my cost of credit. Do the same on the supplier side — am I paying earlier than I need to? If I gave you invoice dates, also calculate DSO, DPO and my cash conversion cycle, showing the working. Finish with the three changes that release the most cash in 90 days. Excel: one tab ranking customers by days late and cash tied up, one tab on supplier payment timing, one summary tab with cash released by each recommended change.
Live version: every payment date is already in the feed, so this becomes measured rather than remembered. "What's my average gap between due date and money landing, per customer, over the last year?"
Watch for: the number to keep is what one day of delay is worth. Once you know a day is worth €4,000, chasing invoices stops being admin and starts being the best-paid hour in your week.
Prompt 6: Month-end variance
What changed, and is it real or a bookkeeping error?
Run this the day after close, before you file the month away.
Month-end variance review. This month by category: [paste] Last month: [paste] Same month last year: [paste] Anything unusual I already know about: [...] Variance table with absolute and percentage change. Flag anything over 15% either way. For each flag, tell me your best guess at which it is: a real change, a timing difference, or a categorization error — and say why. Separately list anything that looks miscoded: a cost in the wrong category, a one-off booked as recurring, a missing accrual. Finish with the three questions I should ask my bookkeeper. Excel variance report, flagged rows highlighted, sorted by absolute change.
Live version: categorization is already consistent month to month, so you're comparing like with like. Inconsistent categories are the single biggest source of false flags in this review.
Watch for: timing differences are noise, miscodes go back to your bookkeeper, and the real changes are usually two or three lines. That's the whole month.
Prompt 7: Project and client profitability
Which work actually makes money once I count the hours?
Revenue tells you who's big, not who's profitable. In most project businesses at least one prominent client is running at or below break-even, and it's rarely the one people assume.
Project and client profitability analysis. Revenue by project or client, last 12 months: [paste] Direct costs booked against each — subcontractors, materials, licences, travel: [paste] Hours or FTE time spent on each: [paste] My fully loaded internal cost per hour: [amount] Overhead I need to cover annually: [amount] Gross margin per project in € and %, ranked. Then contribution margin after charging my internal hours in at cost. Flag anything running below the margin I need to carry overhead, and tell me in € what keeping it costs me. Find scope creep: projects where hours grew but revenue didn't. For the three worst, give me the options — reprice, rescope, change how it's delivered, or exit — each with the annual profit impact attached. Finish with the single project or client that, if fixed, moves my overall margin most. Excel: one tab with the full ranked table, one tab modelling the three worst performers and their options, and a bar chart of contribution margin by project.
Watch for: the answer is almost never "fire the client." It's usually reprice or change the delivery model. Walking into that conversation with the profit impact already calculated makes it a much easier one to have.
Where Claude gets this wrong
Four failure modes worth knowing, because all of them produce answers that look correct.
It fills gaps. Leave out a number and Claude will often supply a reasonable-looking one rather than stop. A forecast with one invented input is indistinguishable from a real one. This is what rule 1 in the context block is for, and it's worth restating in any prompt that matters.
It mixes currencies silently. If you paste transactions in three currencies without saying so, you can get a total that means nothing. Rule 3 forces it to state the rate and date it used, which also makes the number auditable later.
It's confidently wrong about national rules. Anything touching VAT, payment terms legislation or filing deadlines varies by country and changes. Making Claude name the jurisdiction it applied turns a silent assumption into something you can check in thirty seconds.
It answers in chat when you asked for a file. Put the output instruction last and be specific — "build a .xlsx with one tab per scenario," not "make it a spreadsheet." If it hands you a markdown table, ask again; it can do the real file.
Get Claude real numbers in five minutes
Every prompt here works with data you paste in. It's just slower, and slower means you do it less often, which is how a CFO-on-demand quietly becomes a thing you did once.
Zenith connects to 2,400+ EU banks with a single read-only PSD2 connection. That data syncs to Google Sheets, to your own app via our REST API, or straight to Claude through MCP.
Read-only: Claude can see your data, never move money. Data stays in the EU, isn't used to train AI models, isn't shared with third parties.
See how bank sync works → · Check coverage in your country →