Read-Only Budget Anomaly Brief Prompt

Start here

Paste the complete prompt (provided below this guide) into an ordinary AI chat that accepts text. Then answer its questions. You need to supply:

• budget rows or a table with column headers (only the columns that matter is fine);
• which columns mean money in, money out, the date, and any category;
• the time period covered;
• known one-off items you already expect;
• any spending rules you care about (for example a monthly cap per category);
• amount format: currency, decimal separator, and whether negatives mean spending or refunds;
• how you want issues grouped by confidence.

If you leave an input out, the prompt asks for it and lists it under missing inputs instead of guessing.

What it produces

A permission brief with seven sections: role and scope, allowed actions, forbidden actions, anomaly flags, missing or unclear inputs, suggested review order, and a safe handoff message. Each flag cites a row or date from your data, the comparison that triggered it, a confidence level, and a check for a person to make.

How to check the output

• Every flag points to a row, date, or comparable pair you supplied.
• No flag is stated as confirmed.
• The agent is only reading and reporting, never fixing or paying.
• Sensitive identifiers appear as short labels, not full numbers.
• Anything depending on data you did not give is listed under missing inputs.

Fictional shape example: two Utilities entries of 120.00 in the same week could be flagged as medium confidence, asking whether one is a duplicate charge. That is an illustration, not real data.

FAQ

Does it read my spreadsheet for me? No. You paste rows yourself. The prompt explicitly forbids opening, syncing, editing, sending, or scheduling anything.

Can I reuse it next month? Yes. The prompt is written to be copied into an assistant setup and used again; results depend on the data and rules you paste each time.

Verification note

Documented setup instructions only. source reviewed; runtime not tested.

Complete reusable prompt

You are helping me write a read-only permission brief for an AI assistant or agent that will look at a shared household budget spreadsheet and flag possible anomalies for a human to review. The agent must never change the sheet. It only reads what I paste and reports what it noticed and what it could not judge. Do not claim the agent can open, browse, sync, or edit any file, account, or service by itself.

Ask me for any missing inputs first. I need to provide:

• The budget data as pasted rows or a table, with column headers. If it is long, I can paste the columns that matter.
• The columns that define money in and money out, plus the date column and any category column.
• The time period covered.
• Known one-off or unusual items I already expect, so they are not reported as surprises.
• Any spending rules I care about, such as a monthly cap per category or a rule that recurring bills should appear once per month.
• Amount format: currency, decimal separator, and whether negative numbers mean spending or refunds.
• How I want likely issues grouped: high, medium, or low confidence.

Then produce a permission brief with these sections, in this order:

1. Role and scope State that the agent reads only the pasted data for this one review. It does not access the spreadsheet, bank, email, calendar, or any account. It does not edit, rename, move, delete, send, post, or schedule anything. It prepares findings for a human only.

2. Allowed actions List exactly what the agent may do: parse the pasted rows, compare amounts, watch for duplicates, missing recurring entries, category mismatches, dates outside the stated period, blank required cells, and values that conflict with the rules I gave. It may sort and summarise. It may not fill gaps with guesses, change a value to make it fit, or merge rows unless I explicitly allow it.

3. Forbidden actions List what it must not do: no editing the source data, no banking or payment actions, no sending messages, no deleting or archiving anything, no inventing transactions, no treating a suspected error as confirmed, and no exposing full account numbers, card numbers, addresses, or other sensitive identifiers in its output. If those appear in the pasted data, it should refer to them by a short label such as 'row 4 card' rather than repeating the full number.

4. Anomaly flags For each flag, use this structure: row or date reference, what looks unusual, the specific comparison that triggered it, a confidence level of high, medium, or low, and a suggested human check. Keep the reason tied to the data I supplied, not to general financial advice. Do not tell me to invest, borrow, or change providers.

5. Missing or unclear inputs List what the agent could not evaluate because data, a rule, or a column was absent. Say what would make each item checkable. Do not silently skip it.

6. Suggested review order Give a short numbered list of the flags a person should check first, based on the confidence level and the amount at stake. Mark that this is a suggestion, not a decision.

7. Safe handoff End with a short message I can use when passing findings to another household member: it must describe the flags as unconfirmed and ask for a human decision. It must not state that money was moved, corrected, or approved.

Boundaries for the whole brief:

• This is a household budgeting aid, not financial, tax, legal, or investment advice. If a flag could involve fraud, a disputed charge, or a bill I might owe, say that a human should check the original source and contact the relevant provider or institution directly.
• Never state or imply that the agent has performed an action. Use 'report', 'flag', 'suggest a check', and 'prepare for review', not 'fix', 'pay', 'transfer', or 'update'.
• If the pasted data is too incomplete or ambiguous to support even a low-confidence flag, say so and stop rather than guessing.
• Keep the brief compact and practical. I want to copy it into an assistant or agent setup and use it again next month.

Fictional mini example for shape only, not real data: Input rows: date 2025-04-03; description 'Electric'; category Utilities; amount -120.00. Another row 2025-04-07; same description and category; amount -120.00. Example output flag: Medium confidence. Two Utilities entries of 120.00 appear in the same week. Check whether one is a duplicate charge or a legitimate second payment. Do not change either row. Example missing-input note: No monthly cap was supplied for Utilities, so I cannot say whether 240.00 for the month is over budget.

Final check before returning the brief:

• Every flag cites a row, date, or comparable pair from my data.
• No flag is presented as confirmed.
• No action beyond reading and reporting is assigned to the agent.
• No sensitive identifier is repeated in full.
• If any section depends on information I did not provide, it is listed under missing or unclear inputs instead of being invented.

Return only the permission brief and the missing-input list. Do not add general budgeting advice or a product walkthrough.

References and reuse

• ChatGPT release notes

Original TokRepo prompt · CC BY 4.0. Reference documents retain their own rights.