The AI starter kit.

I work at a building materials distributor. I'm giving you two files: our customer list, and a list of new building permits pulled in the counties we serve.

WHAT I WANT
Tell me which permits belong to contractors we already sell to, and which are names we've never done business with.

HOW TO MATCH
Company names will not match exactly. Treat these as the same company:
- Legal suffixes that differ: Inc, LLC, Co, & Sons, Construction, Builders, Contracting
- Abbreviations and misspellings: "Jas. Miller Const" and "James Miller Construction"
- A DBA or a division of a parent we already sell to
- The same address or phone number under a different name, where both files have one

When you are not sure, say so instead of guessing. A wrong match is worse than an unknown one, because nobody checks a match that looks confident.

WHAT TO GIVE ME
Three tables.

1. MATCHED - permit number, the contractor name on the permit, the account it matches, our account number, the rep who owns it, and your confidence (high / medium / low). Sort by permit value, highest first.
2. NEW - permits from contractors with no account. Permit, contractor, address, permit value, type of work.
3. CHECK THESE - anything you matched at low confidence, with the reason, so a person can decide in ten seconds.

RULES
- Every row has to trace back to a line in the files I gave you. Don't add contractors that appear in neither.
- If a field is missing in the source, write "not in file." Don't fill it in.
- No summary paragraph. The tables are the answer.

Here is a list of contractors pulling permits in our area that we have never sold to. I want to know which ones are worth a call.

For each one, search the web and find what you can:
- How long they've been in business, and whether their license is current
- Roughly how many permits they pull a year, and what kind of work
- What they build: single-family, multifamily, commercial, remodel, tenant improvement
- Reviews and reputation, including anything about how they pay their suppliers
- Who to ask for - owner, PM, or whoever actually buys materials
- Whether they look like they already have a supplier we'd have to displace

RULES
- Cite a source for every claim. A link or a publication name.
- If you can't find something, write "couldn't find it." Don't infer it from the company name.
- Flag anything that would make me not want the account: lawsuits, liens, a lapsed license, a pattern of complaints.
- Five lines per contractor, maximum.

Then give me one list at the end: the five I should call first, in order, and the one sentence I'd open with for each.

Here is our customer list with revenue per customer, and this week's new permits.

STEP 1
Work out what a good job looks like for us, using our actual history - the size, the type of work, the geography, and the kind of contractor behind the jobs we make money on. Show me that profile first, in six bullets or fewer, so I can tell you if you've got it wrong before you rank anything.

STEP 2
Rank this week's permits against that profile. For each one:
- A score out of 10
- The one reason it scored that way
- Who should call, if the contractor is already an account
- How urgent it is - do we need to quote this week, or in two months?

Put the top 10 in a table. List the rest underneath in order, one line each.

RULES
- Rank on what's in the data, not on what sounds impressive. A $40K job with a contractor who buys from us every month beats a $2M job with somebody we'll never get on the phone.
- Say what you're uncertain about instead of smoothing it over.
- Don't recommend a permit you can't tie to a line in the files I gave you.

Here are invoices from 10 jobs that were good for us - high revenue, high margin, or both.

Read them and tell me what these jobs actually have in common. Look at:
- What was on them: product mix, categories, number of lines, one delivery or several
- Size: dollar value, square footage, units
- Who bought: type of contractor, how long they've been an account, which rep
- Timing: how long from first quote to final invoice, and whether anything got added mid-job
- Where the margin came from: the mix, the pricing, the volume, or something else

Then write me a one-page profile of our best-fit job: what it looks like, what it's worth, who buys it, and the three signals I could spot early that tell me a job is one of these.

RULES
- Use only what's in the invoices. If 10 jobs aren't enough to support a conclusion, say that instead of reaching for it anyway.
- Give me numbers where you have them - averages, ranges, counts.
- At the end, tell me what you'd want to see that isn't in these files, and what you'd learn from it.

Here is our customer list with revenue per customer by month for the last 24 months, and the permits pulled in our area.

Find me the accounts where something changed:
- An account whose permits are showing up in the permit list, but whose orders with us dropped or stopped
- An account still pulling permits steadily but buying less from us than in the same months last year
- An account that used to buy across several categories and now buys one

For each one: the account, what changed, when it started, and what it's worth at last year's run rate.

RULES
- Rank by dollars at risk, biggest first.
- Normal seasonality is not a signal. If a drop matches the same months last year, say so and leave it out.
- Don't speculate about why. Give me the pattern and I'll make the call.

Here are our last 10 orders and our price list. Audit them the way somebody would who has to answer for the credit memos at the end of the month.

FOR EACH ORDER, CHECK
- Price: does every line match the price list, and the right customer's pricing?
- Quantity and unit: does the unit on the order match the unit we actually sell in?
- Math: extended prices, total, freight, tax
- Products: anything that doesn't belong with the rest of the order, or that's missing a required companion item
- Anything entered twice, or a line that looks like a typo - a quantity an order of magnitude off, a price with a digit missing

WHAT TO GIVE ME
One table across all 10 orders: order number, line, what's wrong, the dollar impact, and how sure you are.

Then two things:
- Total dollars at risk across all 10
- The pattern. Is the same mistake happening more than once? Say which, and where in the process it most likely starts.

RULES
- "No errors found" is a real answer. Don't invent findings to fill the table.
- Show the arithmetic for anything you call a math error.
- Flag anything you can't check without information I didn't give you, and say what you'd need.

Here are our last 10 orders with pricing, and our price list. Tell me where the margin went.

LOOK AT
- Lines priced below list, and by how much
- Discounts applied with no obvious reason
- Freight: what we charged against what the order looks like it cost to deliver
- Small orders and split deliveries - orders that probably cost more to serve than the margin on them
- Add-on lines that were given away
- Any customer here whose margin is consistently below the others

WHAT TO GIVE ME
1. A margin figure per order, with the calculation shown.
2. The three orders where the most margin leaked, what happened, and what it was worth.
3. Three changes we could make, each with the dollars it would have been worth across these 10 orders.

RULES
- If I didn't give you cost data, say so and work in price-against-list terms rather than inventing a cost.
- Don't recommend raising a price on a customer without saying what we'd be risking.
- Every number traces to a line on an order. Show your work.

Here's everything on an order that went wrong: the order itself, the email and text chain around it, and whatever else I've attached.

Work out what actually happened.

GIVE ME
1. A timeline. Every event with a date and who did it: placed, changed, confirmed, picked, shipped, delivered, disputed. Quote the exact words wherever somebody agreed to something or changed something.
2. The point where it went wrong. Be specific - which message, which line, which day.
3. What was actually agreed, as best the documents show, and where the documents disagree with each other.
4. What's recoverable and what isn't.
5. What would have caught this: a step, a check, or a piece of information nobody wrote down.

RULES
- Stick to what's in the documents. Where you're inferring, label it as an inference.
- Where the record is silent on something important, say so. "Nobody wrote it down" is the most useful finding here.
- Don't assign blame to a person by name. Find the gap in the process.