Your role
Finance work has a characteristic that other positions do not have:If you get it wrong, you are liableA product manager can write a requirement incorrectly and fix it in the next version; if finance reports a number incorrectly, it may already have made it into materials for the board of directors.
So this holds from beginning to end:What you can do for them is collect, excerpt, make preliminary calculations, and draft; what you cannot do for them is endorse.
Daily time is spent in three places: finding numbers (several companies' statements, exports from several systems, and descriptions of accounting bases sent by business departments in group chats, scattered everywhere), checking numbers (manually moving figures from one table to another, checking each time you move them), and writing (the hardest part of a monthly analysis to write is the "why", not the numbers).
Context map
| Context | Where | Typical content |
|---|---|---|
| Expense transactions | Approval | Reimbursement forms, payment applications, contract sealing |
| Budget and actuals | Base / spreadsheet | Budget tables, departmental expense ledgers |
| Bases and policies | Cloud documents, knowledge base | Expense standards, accounting bases, financial policies |
| Business explanations | Business group chats, messages mentioning me | Back-and-forth communication on "why did this exceed the limit" |
| Closing rhythm | Calendar, tasks | Monthly closing, quarterly closing, audit timing |
| External data | Requires browser access | Cninfo, exchanges, company annual reports |
The first five rows are in Feishu, and the sixth row is outside Feishu. These two types of data have different credibility requirements:If internal data is wrong, you can trace it back to the source; if external data is wrong, you do not even know at which step it went wrong, so scenario 3 uses a separate set of rules.
Scenario routing
| What the user is saying | Which scenario to use |
|---|---|
| Monthly analysis, business data, why did it fall | Scenario 1 |
| Reimbursement, expense anomalies, approval records | Scenario 2 |
| Annual report, listed company, Cninfo, extract figures | Scenario 3 |
| Budget execution rate, monthly tracking, scheduled | Scenario 4 |
Scenario 1: Monthly business analysis
How it was originally done: Export details from the system, pivot them in Excel, find the accounts with large changes, and then start writing the analysis. When you get to the "reason" column, you get stuck—you know the numbers, but the reason has to be asked of the business side; if you cannot get an answer, you can only write a sentence like "mainly due to the market environment and other comprehensive factors". Writing that sentence on a PPT is equivalent to saying nothing.
Instruction template：
Analyze this monthly operating data; do not restate the numbers. First identify the five changes with the largest amount impact, then distinguish whether they were caused by price, sales volume, product mix, cost, expenses, or one-off factors. For causes that lack supporting evidence, do not make them up on your own; uniformly rewrite them as questions that need to be verified with the business department. If the existing materials are insufficient to reach a conclusion, state what additional materials are needed, and do not fill in numbers based on experience. Finally, output in four parts: management summary, detailed analysis, risk warnings, and next month's recommendations.
Output format：
| Change item | Amount impact | Attribution | Basis |
|---|---|---|---|
| East China revenue decline | −1.84 million yuan | Sales volume decline, price roughly flat (average price −0.8%, sales volume −11.2%) | Detail table "Region-Monthly" sheet, rows 34–58 |
| Marketing expenses increase | +620,000 yuan | Needs verification with the marketing department: In August, two new placements each exceeding 200,000 were added, approval numbers FK-0812 and FK-0827, with the reason field left blank | Approval records, items 112 and 139 |
The most valuable sentence in this instruction：
For causes that lack supporting evidence, uniformly rewrite them as questions that need to be verified with the business department.
Do not let it make up a plausible-sounding reason for you; just have it list "what to ask whom next." Without this sentence, it will fill in every column using industry common sense, looking complete but actually all guesswork.
Acceptance checklist：
- Can the amounts of the five changes be recalculated from the detail?
- Does each cause have a basis, and have all unsupported ones been converted into questions?
- Do the risk warnings sneak in its own judgments?
Scenario 2: Filter anomalies from approval records
When reviewing expenses at month-end, summing them up is not hard; the hard part is picking out the few entries that "seem off." Travel over standard, the same invoice reimbursed twice, or a department's entertainment expenses suddenly doubling are hidden among hundreds of approval records and are very hard to spot by human scanning.
Instruction template：
Retrieve all approved reimbursement forms from last month's Feishu approvals, summarize by department and expense type, and generate one detail table and one summary table. Then, according to the following criteria, filter out the documents that need my review: single amounts exceeding (amount); two or more submitted by the same person on the same day; identical invoice numbers or amounts; and obvious mismatches between expense type and description. List the filtered items in a separate table, indicating the suspicious reason and the corresponding approval link. Do not conclude that there is a problem; only say why a person needs to take a look.
Acceptance checklist：
- Spot-check a few entries it did not filter out to see whether the rules missed anything
- See how many false positives are in the filtered results; a false-positive rate that is too high means the threshold was set too loose
Negative clause that must be retained：
Do not conclude that there is a problem; only say why a person needs to take a look.
This step only produces a review list; "this entry has a problem" can only be said by a person. This requirement is not a formality; its effect is to prevent the list from being sent out directly as a conclusion.
Scenario 3: Extract figures from a listed company's annual report
Before handing off this task, the instruction must clearly state seven things; missing even one raises the review cost later by a notch:
| # | Must specify | Example |
|---|---|---|
| 1 | Where to look | Annual reports officially disclosed on Cninfo, exchanges, or by the company |
| 2 | Which company | Write the full company name, don't use an abbreviation |
| 3 | Which period | Annual, semi-annual, or a specific month |
| 4 | Which metrics | List them all at once, don't add them in several batches |
| 5 | What unit to use | Yuan, ten-thousand yuan, or one hundred million yuan |
| 6 | What formula to use | Specify the numerator, denominator, and period basis |
| 7 | What output format | Excel, number of decimal places, thousands separator, checking requirements |
Instruction template：
I am a financial analyst and need to extract the financial data of (full company name) for (year) from the financial statements in that year's annual report on Cninfo. Use ten-thousand yuan as the unit for all data, keep two decimal places, use thousands separators, output in Excel, and keep only the metrics and their corresponding amounts. Metric list: (list them all at once). Calculate ratios according to the formula basis I have confirmed, make sure there are no error values in the table, and check it yourself before output.
Ninety-five percent is not equal to one hundred percent. Someone checked its output item by item against manually compiled results, and the accuracy rate was around ninety-five percent. Work that used to take two days, including checking, can be finished in over ten minutes. But the remaining half a percent is enough to be fatal: period misalignment, inconsistent bases, and ambiguous formula parentheses are all common errors. Metrics involving average balances, such as turnover days, are especially prone to problems.
Hard requirements：
Final data must be traced back to the original disclosure.
This step cannot be skipped. You can not know a number, but you cannot report that number incorrectly with great confidence.
Scenario 4: Monthly tracking of budget execution rate
Instruction template：
Every month on the 3rd at 9 a.m., summarize last month's expense ledger multidimensional table by department, compare it with the budget table, and generate an execution rate table: budget amount, actual amount, execution rate, variance amount. Departments whose execution rate is above (threshold) or below (threshold) should be marked separately. Save the results to (Feishu folder), and also push the list of overspending departments to me. If the data source has not been updated or fails to read, remind me; do not make up a report with old data.
That last sentence is the life-saving talisman for this type of task. Automation will repeat a single correct or incorrect action many times. Once the upstream table has not been updated, it silently uses last month's data to generate a report that looks normal, which is far more dangerous than reporting an error directly.
Run it manually once before scheduling automation. After confirming that the table structure, thresholds, and summary basis are all correct, then set it as a recurring task.
Basis first
Financial errors are often not in the calculation, but inbasis drift: this time calculated tax-inclusive, last time calculated tax-exclusive.
For any task involving calculation, have it restate the basis before starting:
Before calculating (metric), first restate the formula you intend to use, the numerator and denominator basis, and the period range, and wait for my confirmation before calculating.
This step takes ten seconds and can save an entire recalculation. What most deserves to be consolidated into a custom skill is your company's basis library: expense classification standards, calculation formulas for each ratio, and the correspondence between report accounts.
Traceability requirements
Financial outputs require more traceability than those of other roles — audits need to check them, reviews need to consult them, and next year's budgeting needs to refer to them.
For every piece of material produced with AI participation,the source, basis, and reviewer must be written in the documentExcerpted external data is archived together with the source link and excerpt date, so the next time someone questions a figure, it can be traced all the way back to the original annual report.
These traces may seem cumbersome, but they are the only evidence if a problem arises.You must be able to prove where this figure came from.
Fallback Strategy
Incomplete data source: State plainly which days and which line items are missing; do not fill in with prior-period data or averages.
Definition document cannot be found: First ask the user for the definition, restate it to confirm before producing output; do not default to industry conventions.
Annual report page structure changed and cannot be scraped: Explain which step it got stuck at; do not pass off second-hand data found via search as the original text of the annual report.
Upstream and Downstream Connections
- Upstream：
job-ops(activity expenses),job-sales(quotes and discounts) - Downstream: The to-be-reviewed checklist is created as Feishu tasks and assigned to the corresponding department contacts
Honesty Boundaries
- What it can do is collect, excerpt, do preliminary calculations, and draft; what it cannot do is endorse
- Any external data and report conclusions can only be used after manual review
- "There is a problem with this item" can only be said by a human; this Skill only produces a to-be-reviewed checklist