Project: Build a Personal Finance Tracker with AI Prompts

In this project you will build a personal finance tracker driven by a chain of prompts — automatic transaction tagging, a monthly summary, anomaly spotting, and honest, non-generic advice tied to your real numbers. The deliverable is a small workflow you can run every month on a CSV exported from your bank.

1. Introduction

Most personal-finance apps stop at colourful pie charts. The hard work — interpreting the chart, deciding what to change, sticking with it — is left to you. AI is genuinely useful here, because the bottleneck is not the data; it is the thinking on top of the data. This project gives the thinking a structure.

You will need one CSV: a 3–6 month export of your bank transactions. Most banks let you download this. Anonymise it if you like.

Important: This project produces personal reflection prompts and pattern analysis — not regulated financial advice. Treat the outputs as a structured conversation with yourself, not as guidance from a financial adviser. Be careful what you upload to any third-party AI tool; redact account numbers and run the prompts on a tool whose privacy terms you have read.

2. The Concept Explained

The chain has four steps: clean and tag the transactions, summarise the month, spot anomalies, and produce a tailored "next-month plan". Each step has a focused prompt with a clear input and a clear output.

Four focused prompts turn a raw transaction list into a reflective monthly plan.

3. The Problem Without Structure

Vague finance prompt

Here's my spending. How can I save money?

The model returns advice that could apply to anyone: "cancel unused subscriptions", "cook more at home", "make a budget". None of it touches your actual numbers. Generic advice is forgettable, and forgettable advice is unactioned advice.

4. The Solution: A Four-Prompt Reflection Chain

Step 1 — Tagging prompt

You are a careful personal-finance assistant.

Tag each transaction with ONE label from this fixed list:
[Rent, Utilities, Groceries, Eating out, Transport, Subscriptions,
Health, Personal care, Clothing, Entertainment, Travel, Gifts,
Savings/Investments, Income, Transfer, Fees, Other]

Rules:
- Choose exactly one tag per row.
- If a merchant is ambiguous, prefer "Other" and add a note.
- Do NOT invent transactions or change amounts.
- For income rows, use "Income" regardless of merchant.
- Use the merchant string, the amount sign, and any reference field
  as clues.

Return a CSV with the original columns + a new "tag" column and a
"confidence" column (high / medium / low).

Transactions:
"""
date, merchant, amount, reference
2026-04-02, "TESCO 4421 BRISTOL",  -38.42, ""
2026-04-02, "TFL TRAVEL CH", -3.20, ""
2026-04-02, "SPOTIFY UK", -9.99, "monthly"
... (paste 3–6 months of rows)
"""

Spot-check the rows tagged "low" confidence and correct them. A correctly tagged dataset is what makes every later step work.

Step 2 — Monthly summary prompt

Using the tagged data, summarise each month.

For each month return a Markdown block with:
- Total income
- Total spending
- Net (income − spending)
- A breakdown table: tag, amount, % of total spending
- The top 5 individual transactions (merchant + amount)
- One paragraph: "what stands out this month vs. the previous one"

Tone: neutral, factual, no moralising. Use my currency from the
data. Round to whole units, not pence.

Step 3 — Anomaly prompt

You are looking for ANOMALIES, not patterns.

Given the per-month summaries above, list:
- Any tag that increased by > 25% vs. the rolling 3-month average
- Any single transaction larger than 2x the median transaction in
  that tag
- Any new merchant appearing 3+ times in a month (possible new
  subscription or habit)
- Any month where income/spending diverged unusually

For each anomaly return:
- Month + tag + amount
- A one-sentence explanation of why it's flagged
- A possible benign cause (so I don't panic) and a possible
  concerning cause

Do NOT moralise. Do NOT give advice yet. This step is observation
only.

Separating observation from advice is essential. Finance prompts that conflate the two leap to "stop drinking lattes" before they have understood the data.

Step 4 — Next-month plan prompt

You are a thoughtful friend who happens to be good with money,
not a financial adviser.

Based ONLY on the summaries and anomalies above, propose 3
concrete actions for next month.

Rules:
- Each action must reference my actual numbers (e.g. "Eating-out
  averaged £312/month over the last 3 months — try capping at
  £220 by limiting weekday lunches out to 2/week").
- No generic advice ("make a budget", "cancel subscriptions").
- No moralising language ("you should", "you wasted").
- For each action, include: the trigger (which data point), the
  experiment (what to try for one month), the success metric (how
  you'll know it worked), and a "would-I-actually-do-this?" honesty
  check (1 sentence).
- End with one question I should ask myself before next month
  starts.

I want pattern-spotting and decision support, NOT regulated
financial advice. If I asked anything that needs a professional,
say so clearly.

Sample action: "Trigger: Subscriptions jumped from £37 to £62 last month — three new services in March that you haven't used twice. Experiment: cancel two of them for April and re-subscribe only if you reach for them. Success metric: subscriptions back under £45 next month. Honesty check: realistic — you cancelled Apple Arcade in January and never missed it."

5. Step-by-Step Breakdown

Use a fixed tag list. Letting the AI invent tags produces 50 inconsistent ones. A fixed list of 15 covers 95% of real spending.
Tag with confidence labels. The "low confidence" subset is short — review it by hand. Everything else can be trusted.
Separate observation from advice. Step 2 and 3 describe what happened. Step 4 decides what to do. Mixing them produces lazy advice.
Tie every action to a real number. "Eat out less" is generic. "Cap eating-out at £220 — you averaged £312" is specific. The prompt forces this.
Include an honesty check. Asking the model "would I actually do this?" produces realistic suggestions, not idealised ones.
Re-run monthly. The chain is most valuable when run consistently. A folder of twelve monthly reports tells you more about yourself than any app's dashboard.

Tip: Add a "wins" prompt at the end: "Identify two things I did well this month — tags I improved on, savings I hit, an anomaly I caught early." Positive reinforcement keeps the habit alive longer than guilt does.

6. Practice Exercises

Exercise 1

Export the last three months of your bank transactions. Run only step 1 (tagging) and step 2 (summary). Read the breakdown. You will probably learn one thing about your spending that surprises you — that single learning is worth the project.

Exercise 2

Run step 4 with and without the honesty check rule. Compare the two plans. The honest one is shorter and more boring — and far more likely to actually happen.

Exercise 3

Set a recurring monthly reminder to run the chain on the previous month's CSV. After three months, ask the model: "Looking at my last three plans and what actually happened, what does the data say about which plans I follow through on?" Meta-analysis is where this project becomes a habit.

7. Key Takeaways

The bottleneck in personal finance is reflection, not data — AI helps with the reflection.
Use a fixed tag list. Confidence labels make manual correction fast.
Separate observation from advice. Generic advice happens when the two are mixed.
Tie every recommended action to a real number from your data — and an honesty check.
This is reflection support, not regulated advice. Use a professional for anything beyond pattern-spotting.

Discussion

Project: Create an E-Learning Course Outline with AI Project: AI-Powered News Summarizer Using RAG Concepts