Claude Finance Kit: 50 Prompts, 152 Skills and Agents

How we chose these 50 Claude finance prompts

The 50 prompts below were selected and rewritten against five criteria, applied to real finance tasks rather than demonstrations.

• Tested against a real task, not a demo. Every prompt below was run against an actual budget vs actual table, a CIM excerpt, a comps set or a board pack draft, not a hypothetical one.
• One paste, one usable draft. A prompt that needs three follow-up messages to produce something usable failed the bar for this page; that job belongs to a skill, not a prompt.
• No proprietary templates. Nothing here assumes your firm's specific chart of accounts, covenant definitions or LP reporting format; those live in the skill packs linked below, where they belong.
• What we ignored: which model answers best. Leaderboard position changes every few weeks. Every prompt here is written to work across current frontier models, not to flatter one benchmark run.
• What we corrected: the count. An earlier version of this page promised 500 prompts and 150 skills. Neither number traced to a file anyone could open. The 50 prompts below are all here in full; the 152 skills and agents referenced later are counted from Oria's own published packs, not rounded up.

Key takeaways

The stakes for getting this right are real. Gartner's November 2025 survey of 183 CFOs and senior finance leaders found 59% of finance functions now use AI, up from 58% a year earlier, with Marco Steecker, senior director of research in Gartner's finance practice, noting that "the momentum in finance AI adoption has slowed following a sharp increase from 37% in 2023 to 58% adoption last year." Adoption has plateaued at the exact moment teams are deciding what to standardise. The five rules below are what that decision comes down to.

• Start with the prompt that matches a document already open. A variance table wants the variance prompt; a data room wants the CIM prompt. Matching the prompt to the artifact you already have is most of the work.
• A prompt is only as good as what you paste into it. Vague inputs, a screenshot with no numbers, an unlabelled tab, produce vague output no matter how carefully the prompt is worded.
• A prompt has no memory of your conventions between chats. The moment the same method needs to run the same way every month, on the same chart of accounts, you have outgrown a prompt and need a written skill.
• Two people running the identical prompt can still get back two different structures. A written skill collapses that variance; a bare prompt, run by different people, will not.
• Every function below has a purpose-built pack behind it. 152 skills and agents across four packs exist for the moment copy-pasting a prompt each month stops being the efficient choice.

The Claude finance prompts compared

The table applies the five rules above to the five finance functions the prompts cover. The columns are the five takeaways, in the same order, so a row tells you what to try first, what to paste in, where it stops, whether it repeats, and which pack to read next.

Claude Prompts for FP&A: What Should You Ask For First?

Start with the variance narrative prompt, because every finance team already has a budget vs actual table sitting somewhere. Paste it in whole, ask for the three biggest drivers of the gap, and a board-ready paragraph comes back in seconds. BILL's 2026 State of AI in Finance report, surveying 500 finance leaders at small and midsize companies, found that AI tools are saving financial operations employees an average of 21 hours a week; a well-matched prompt is where most of that time comes from.

• List every line in this P&L that moved more than 10% versus last month, and say why in one sentence each.
• Turn this cost centre report into a one-page summary a non-finance manager could act on.
• Check this headcount plan against the approved budget and flag every role added without a matching line.
• Convert these five KPI definitions into a single glossary paragraph the whole team can cite.
• Draft three questions a CFO would ask about this forecast before approving it.
• Summarise this quarter's opex trend in four bullet points, oldest to newest.
• Rewrite this budget memo assuming the reader has thirty seconds, not five minutes.
• Compare these two forecast versions and list only what changed between them.

The trade-off: a prompt like this hands back a draft, not a decision. It has no memory of your chart of accounts between chats, so the same phrasing can catch a different set of drivers next month depending on how the data happens to be laid out.

Claude Prompts for Treasury and Cash: Which One Catches a Covenant Breach?

Claude can carry real modelling work underneath a prompt like this. FundamentalLabs, testing Claude Opus 4 as an Excel agent, reported that it passed five of seven levels of the Financial Modeling World Cup and scored 83% accuracy on complex Excel tasks, according to Anthropic. Good enough to draft a schedule, not good enough to release one without a named person checking it first.

• Rank these five bank accounts by idle cash balance and suggest a sweep order.
• Flag every payment in this run over $50,000 that lacks a matching purchase order.
• Summarise this quarter's FX exposure by currency in one table description.
• Turn this debt schedule into a plain-English maturity wall, year by year.
• List the three assumptions in this cash forecast most likely to be wrong.
• Draft a one-paragraph liquidity update for a lender who reads only the headline.
• Check this intercompany loan schedule for interest rates outside the approved range.
• Explain this covenant definition in one sentence a non-finance manager would understand.

The trade-off: treasury prompts are only as current as the data you paste in. A 13-week cash forecast built from last week's export is confidently wrong, and the prompt has no way of knowing the export is stale.

Claude Prompts for Valuation and Comps: How Do You Keep the Peer Set Honest?

Vals AI's Finance Agent v2 benchmark, last updated 15 September 2026 across 927 expert-reviewed questions, has the leading model at 60.60% accuracy and Claude Opus 5 at 58.63%. A comps prompt inherits that error rate, so treat the peer set it returns as a first cut, not a final answer.

• Score this list of 20 potential comps on business-model similarity to the target, one to five.
• Explain why this precedent transaction might not be comparable to the deal in front of you.
• Bridge this enterprise value to an equity value per share, showing every step.
• Stress-test this valuation range against a 2-point change in the exit multiple.
• List the three most aggressive assumptions in this management case.
• Summarise this fairness opinion in three sentences a board member would actually read.
• Check this football field chart for a method whose range looks too narrow.
• Draft the one-paragraph valuation summary for the front of the deck.

The trade-off: a comps prompt will politely accept whatever peer list you give it. It has no independent view of whether your 25-name long list is really a 6-name peer set with 19 names hiding the analysis, which is a judgement call, not a paste.

Claude Prompts for Deal Diligence: What Can a Single Prompt Actually Flag?

The Finance Agent Benchmark paper (arXiv 2508.00828) put 537 expert-authored questions to frontier models with live search and EDGAR access. The best performer, OpenAI's o3, reached 46.8% accuracy at an average cost of $3.79 per query, and no model crossed 50%. A diligence prompt is a useful first pass, not a substitute for a second reader.

• Summarise this customer concentration table in one sentence a partner would remember.
• List every related-party transaction mentioned across these three documents.
• Explain this EBITDA adjustment in plain language and say whether it looks aggressive.
• Draft five diligence questions this management presentation raises but does not answer.
• Compare this quality of earnings summary against the CIM's own EBITDA bridge.
• Flag any inconsistency between this deck's growth story and the historical financials.
• Summarise the key person risk in this org chart in two sentences.
• List every representation in this SPA that depends on a fact not yet verified.

The trade-off: a diligence prompt reads one document at a time. The red flags that actually kill deals usually sit in the gap between two documents that individually look fine, and catching that gap needs the cross-document memory a single prompt does not have.

Claude Prompts for Board and Investor Reporting: Who Reads the Draft First?

These two prompts draft the sections a controller usually writes last, under the most time pressure, so they are worth trying even if nothing else on this page is.

• Rewrite this KPI dashboard commentary so a first-time reader understands it in one pass.
• Draft a one-line headline for each of these six KPIs, worst news first.
• Check this investor letter for any number that appears twice with two different values.
• Summarise this quarter's risk register into the three items a board would ask about.
• Turn this bullet list of updates into three flowing paragraphs, board-pack style.
• Draft the single most important sentence this board pack needs on page one.
• Compare this quarter's outlook language against last quarter's for any quiet walk-back.
• List every acronym in this deck that a new board member would not recognise.

The trade-off: the output reads like your board pack, not like the CFO who signs it. The tone, the acronyms and the level of hedging on a bad quarter are things you will still edit by hand, and that editing is the part a prompt cannot shortcut.