2. The Six Questions You Have to Ask Out Loud

The empty space in an honest brief

The research brief from the earlier project is a good document — every claim numbered to a source, every figure verbatim, silence declared where the sources are silent. It also contains no evidence that a single electrician would hand over money. That is not a fault in the document or in the tool. Twelve packet items, invented and labelled as such, would still only record what people said in the past, to someone else, about something else. Demand is what a person does when a priced offer arrives in a busy week.

So the useful part is the closing list of six unknowns: enquiry volume, the share that goes unanswered, average job value, whether owners believe they lose work, who answers the phone, and whether anyone has ever paid for help. None can be found by better searching. Each lives inside one business, known to one person. The list is an agenda.

Getting a tradesperson to talk

Solo electricians are unreachable during work hours and delete anything that smells of sales. The five-part ask from Rob Fitzpatrick's approach to customer conversations works because it gives no reason for suspicion: state the subject, state the stage, admit ignorance, put them on the pedestal they deserve, and box the time at ten or fifteen minutes. No product description — the polite answer to a pitch is encouragement, and encouragement is worth less than nothing. Trade counters at wholesalers between half six and eight, and text messages rather than email, do most of the work.

Then the discipline: ask about specific past events, not opinions or predictions. An invented owner, Dave, shows both versions — the leading question that produces "it's a nightmare," and the sequence about last Thursday that yields a date, a count, a three-hour callback latency, a lost job and a £200–250 value. Hedges and uneven detail are the tell that a memory is real.

The sixth unknown sorts people into three positions: already paying an answering service or lead platform, spending themselves through workarounds, or doing nothing at all. Nothing is a finding too. Saturation, not sample size, tells you when to stop — usually ten to fifteen within one narrow group.

Filing so five conversations compare

A fixed note schema goes in the project instructions, with unmentioned items written as not reported. Filename prefixes keep the fabricated twelve apart from real notes, because nothing a chat produces saves itself and retrieval over a grown knowledge base must tell them apart by name. Then rerun the brief — and read hardest for where a real answer contradicts the packet.

The brief that came out of that project is a good document. Every claim in it points at a numbered source, every number appears verbatim somewhere in the packet, and where the sources are silent it says so. It is honest. It is also, on the question that matters, empty.

Nothing in it proves that a single electrician would hand over money.

That is not a flaw in the brief and it is not a flaw in Claude. It is a fact about what research packets are. The twelve items in that packet are invented and labeled as invented, but even if all twelve were real interview notes collected by somebody else, they would still only be a record of things people said in the past to someone else about something else. Demand is not a fact stored in a document. Demand is what a particular person does when a particular offer is put in front of them, at a price, in a week when they are busy. The nearest thing you can get to it before you have an offer is a careful account of what people already do and what they already pay for. That account does not exist yet. It exists in the heads of people currently standing in someone's loft with a torch in their mouth.

So the brief's real value is not its conclusions. Its real value is that last section, the one the rewritten project instructions forced it to produce: the list of six things marked unknown. Enquiry volume. What share goes unanswered. Average job value. Whether owners believe they lose work because of it. Who currently answers the phone. And whether anyone has ever paid for help with any of this.

Look at those six for a moment and notice what they have in common. Not one of them can be found by better searching. Each is a fact that lives inside one specific business and is known to one specific person. How many calls came in yesterday while you were on a job is not published anywhere. Whether your wife picks up the phone in the afternoon is not published anywhere. Whether you once paid an answering service sixty pounds a month and cancelled it after two months is not published anywhere. An agent with a search tool cannot get at any of it, and an agent without one certainly cannot. That is precisely why the honest brief marked them unknown instead of filling them in, and it is why the unknowns section is more useful than the rest of the document. It is an agenda. Six things to ask out loud.

Which brings up the part that has nothing to do with software at all. You have to get a tradesperson to talk to you.

This is harder than it sounds and easier than people fear. Harder, because these are the least reachable customers in the world during working hours. A solo electrician is up a ladder, or driving, or has his hands inside a consumer unit. He is not reading email. He is not on a video call. Whatever administrative life he has happens in the evening, after dinner, badly, and he resents it. Anything that smells like a sales approach gets deleted or hung up on within about two seconds, because he gets a dozen of them a week from lead-generation companies.

Easier, because most people like being asked about their work by someone who obviously does not know how to do it.

That second point is the whole basis of the approach. Rob Fitzpatrick's book on customer conversations, The Mom Test, lays out a way of asking for time that works because it removes any reason to be suspicious of you. It has five parts, and they go in order. You say what you are looking into, specifically: how independent trade businesses handle calls and enquiries that come in while the owner is out on jobs. You say what stage you are at: early research, nothing to sell, no product exists. You admit your weakness: you have never run a trade business and you do not know how the day actually works. You put them on a pedestal, which sounds manipulative and is not, because it is true: they are the person who does this every day and you are the person guessing. And then you make a small, specific ask with a hard boundary on it. Ten minutes. Fifteen at most.

Notice what is absent. There is no description of a product. There is no "I'm building something that automatically texts your missed calls." The moment you say that, you have changed what the conversation is. You have made it a pitch, and the polite response to a pitch is encouragement. He will tell you it sounds handy. He will mean nothing by it. You will write it down as evidence, and it will be worth less than nothing, because now you believe something false.

The request must not promise a product. That is not modesty. It is the only way to keep the answers usable.

Where do you actually find these people? Two channels do most of the work. The first is physical and slightly unglamorous: trade counters at wholesalers, early. Between about half six and eight in the morning, a stream of self-employed plumbers and electricians walks in to pick up parts for the day. They are stationary, briefly, and waiting. This is a five-minute conversation, not a fifteen-minute one, and you should treat it that way. Ask one or two things, thank them, leave.

The second is text messages. Not email, which they do not read. A working contractor will answer a text far more reliably, because a text is how customers reach them. So the message reads roughly like this: Hi Dave, I'm doing some independent research into how solo and small-crew electricians handle incoming calls and quotes while they're out on jobs. Not selling anything at all, just trying to understand how it actually works. Could I grab ten minutes by phone this week while you're driving between jobs, or buy you a coffee?

That message does all five parts in four sentences. It names the subject, states the stage, admits ignorance, respects their expertise by asking to learn from it, and boxes the time. And it offers the phone while driving, because that is dead time they already have and they know it.

Now the conversation itself, which is where most people waste the access they worked so hard to get.

The rule is simple to state and unnatural to follow: ask about specific things that already happened, not about opinions, feelings, or what someone would do. Opinions are free. Predictions about your own future behaviour are free too, and worse, they are flattering to make. If you ask a man whether missing calls is a problem, he will say yes, because saying no would mean admitting he runs a sloppy business. If you ask him whether he would use something that texted people back automatically, he will say that sounds useful, because disagreeing with you would be rude and agreeing costs him nothing. Neither answer tells you anything about his business. Both feel like progress. That is the trap.

The past is different. The past is fixed. He cannot flatter you with it, because it already happened, and if you ask him for enough detail he will start telling you the story rather than managing your feelings.

Here is what that looks like in practice. I am going to use an invented owner for this, because I have to: I have no real transcript, and inventing one and dressing it up as real is exactly the sin the last chapter was about. So: Dave, invented, labeled invented, a solo domestic electrician with one van, mostly rewires and fault-finding, no office and no one answering the phone. Every word he says here is written by me to show the shape of the exchange, and none of it is evidence about real electricians.

The bad version goes like this.

You ask: Don't you hate it when you miss customer calls in the middle of a job?

Dave says: Yeah, it's a nightmare, you're losing work all the time.

You write down: owners strongly aware of lost work from missed calls. Confirms the packet.

That exchange is worthless, and it is worth understanding exactly why, because it looks like a hit. The question told him the answer. It presumed the emotion, it presumed the loss, and it invited him to agree. He agreed. He also would have agreed if he answers every call on speaker and has never lost a job in his life, because "it's a nightmare" is what you say about a thing someone has just described as a nightmare. You learned nothing about his triage behaviour and nothing about money. And then the second bad question makes it worse: would an automatic text-back service save you time? Yeah, probably, he says. You have now got two yeses and no facts.

Same ground, reopened properly.

You ask: Think about last week. When you were on a job and the phone rang, what actually happened?

Dave says: Depends. If I'm on the floor I'll take it. If I'm up in a loft or I've got the board off, it rings out.

You ask: When was the last time one rang out on you?

Dave says: Thursday, I think. Two of them Thursday afternoon.

You ask: What did you do about those two?

Dave says: Rang them back about six, when I got in the van. One didn't answer. The other one, the lady, she'd already had someone else out.

You ask: How long between her call and yours?

Dave says: Three hours? Bit more.

You ask: Do you know what the job would have been worth?

Dave says: She said it was a socket sparking in the kitchen, so, fault-find and probably a new circuit. Two hundred, two-fifty, could've been more if the wiring was bad.

Read those two exchanges next to each other. In the second one you have a date, a count, a specific triage rule with the physical reason behind it, a real latency measured in hours, a realised outcome where a competitor got the job, and a job value in his own numbers rather than yours. You did not ask him what he thinks his missed calls cost him per month, which would have produced a made-up figure he would then defend. You asked what happened to one job, and he told you, because he remembered it.

Four of the six unknowns just moved. Not answered, but moved: you have one data point on volume, one on the unanswered share, one on job value, and a direct account of who answers, which is nobody. And the tell that this is real material rather than politeness is in the texture. He hedged. He said "Thursday, I think" and "three hours, bit more." He gave you a range on the money and a reason the range exists. Remembered answers come out uneven, with corrections and irrelevant detail attached. Polite answers come out smooth, general, and in the present tense: it's a nightmare, you're always losing work, I'd definitely use something like that. When you hear a clean confident generalisation, you have probably just been agreed with. Ask when it last happened, and either the story arrives or the confidence evaporates. Both of those are useful.

Which leaves the sixth unknown, and it is the one that decides whether any of this becomes a business.

Has he ever paid for help with this?

Do not ask it as a hypothetical. Would you pay for someone to follow up your leads is the same broken question as would you use an app, and it produces the same courteous fog. Ask instead: what have you tried or paid for, ever, to deal with incoming calls or quotes? Answering service, receptionist, apps, one of the lead sites, your wife, anything?

The answer sorts him into one of three positions, and the difference between them is enormous.

The strongest position is that he already pays money at this problem. A live answering service. A virtual receptionist. A subscription to a lead platform like Checkatrade or Angi. A field-service tool such as Jobber or Housecall Pro with a call-handling add-on. If he pays, the argument about whether the problem is real is over. Budget already exists and has already been approved by the only person who has to approve it. Your remaining job is to find out where the thing he pays for lets him down: too slow, too expensive per lead, the receptionist does not know what a consumer unit is, the leads are shared with four other electricians.

The middle position is that he pays no money but spends himself. He pulls the van over to answer. He set up an auto-reply on his phone. His partner monitors the line in the afternoon for nothing. He does an hour of callbacks every evening on the sofa. No invoice exists, but he has rearranged his behaviour and his household around the problem, and that is a strong signal. People do not build workarounds for things that do not hurt.

The disqualifying position is the quiet one. He misses calls, the voicemail is full, he does not ring back, unaccepted quotes go silent and stay silent, and he takes no action of any kind. He may still tell you missed calls cost him a fortune. His behaviour says otherwise. Usually the reason is simple and unglamorous: he already has more work than he can do, most of it from referrals, and losing a stranger's call costs him nothing he can feel. A problem with no spending against it and no workaround around it is dormant. You can build a perfect solution to it and he will not buy, because he was never trying to solve it.

That distinction is worth more than any statistic you could have found in a packet. Stated willingness to pay is conversation. Existing spend is evidence. A workaround is evidence. Nothing is a finding too.

How many of these do you need? Fewer than you would think for a signal, more than you would like for confidence. This is pattern matching, not statistics, so there is no sample size that makes it valid. What you are looking for is the point where new conversations stop teaching you anything, which people who do this work call saturation: you can predict what the next person will say about their triage habits and their objections before they finish the sentence, and independent people start describing the identical mechanism without prompting. Several plumbers telling you separately that emergency callers book whoever answers within about five minutes, which makes an evening callback pointless, is saturation arriving. For one narrow group, one trade, one or two vans, no office manager, that usually lands somewhere around ten to fifteen conversations. Three or four is a signal worth acting on cautiously. It is not proof, and you should say so out loud in your own notes so that you do not quietly forget it later.

And if you get to the tenth conversation and people are still describing completely different worlds, that is not bad luck. It means your definition of who you are studying is too wide. Emergency domestic repair and planned commercial subcontracting are not the same business with different vans. Split them, and run six to ten in each.

Getting five conversations into one comparable pile

All of that is worthless if it lives in your head and in four different notes apps.

The workspace from last time is already the right place to put it. The Claude project you built, with the numbered packet in its knowledge base and the rewritten instructions demanding a source for every claim, was never only for the packet. It is a place where sources accumulate. Each conversation becomes a new numbered source in it, sitting alongside the invented twelve, and the whole point of the exercise is that the brief can then be run again against a knowledge base that has real material in it.

Two pieces of setup make that possible, and the reason for each is worth understanding before you build it.

The first is a fixed note structure. Not a template you keep in your head, but a schema written into the project instructions, so that every transcript you paste gets turned into the same shape. Something like: participant identifier and trade profile, meaning trade, crew size and typical jobs; what they do when the phone rings mid-job; what they have paid for or built as a workaround; how they handle quotes that go quiet; and a short list of verbatim quotes describing specific past events. Then one hard rule: anything the person did not mention is written as not reported, never inferred, never filled in.

That last rule is the same discipline that caught the invented forty percent figure, applied one layer earlier. And the fixed headings are what let you compare. Five conversations written into identical fields can be lined up field by field, and the pattern in the middle column is visible. Five loose summaries cannot be compared at all, because each one emphasises whatever struck you that evening, and the thing Dave did not mention looks exactly like the thing you forgot to ask him. A blank in a structured note is information. A gap in a paragraph is invisible.

Now, project instructions live in the project's set-instructions field and get prepended to every conversation inside the project, which is why they are the right home for a schema and the wrong home for a one-off task. They cost context on every single turn. Keep them to the role, the schema and the output rules.

The second piece of setup is the boundary between invented material and real material, and this one matters more than it looks. Your knowledge base is about to contain twelve fabricated source items and a handful of genuine conversations. If nothing marks them apart, a later synthesis will blend them, and you will end up with a brief that cites Dave and the invention in the same breath. So mark them mechanically, in the filenames themselves. Anything synthetic or structural gets a prefix that says so, mock or template. Real notes get a participant-numbered interview-notes name. Then add a data-separation rule to the project instructions: files with those prefixes are structural references only, never to be quoted, summarised, or attributed as findings, and synthesis must draw only on the interview-notes files. Put a matching marker line inside each synthetic document too, declaring it a synthetic reference, so the labeling survives being read out of context.

The filing itself is manual, and you should know that going in, because it is a common surprise. Chats inside a Claude project do not share memory with each other, and nothing a chat produces is saved into the knowledge base by itself. The only things that persist across conversations are the project instructions and the files you upload. So the loop for each interview runs like this: open a fresh chat in the project, paste the rough transcript, ask for it in the standard note format from the project instructions, read the output and check it against what you remember, save it locally under its participant number, then upload it through the add-content control in the knowledge panel on the right. Individual files can run up to thirty megabytes and can be plain text, Markdown, PDF, Word or CSV, so notes are never going to strain it. Note that a PDF added to project knowledge is handled as extracted text, which is a reason to keep notes as plain text or Markdown rather than as scans of anything.

Then, once several are in, open another fresh chat and ask for cross-case analysis: scan the interview-notes files, build a comparison of each participant against their triage action, their existing spend and their quote follow-up process, name the patterns shared by three or more of them, flag where trades disagree with each other, exclude the mock and template files, and cite every finding by its filename. On a paid plan, once the knowledge base grows past what fits directly in context, Claude switches to retrieving relevant chunks with an internal knowledge-search tool rather than loading everything at once, which is fine for this and is another reason the filename convention is doing real work: the retrieval has to be able to tell your fabrications from your findings by name alone.

The last step is the one to actually care about. Run the original brief again, same five-part request, against the enlarged knowledge base, and read it for three things. Which of the six unknowns now have real answers behind them. Which are still silent, because you kept forgetting to ask. And which real answers contradict the invented packet.

That third one is the prize. A contradiction means the world just corrected your document, which is the only direction correction can usefully travel. If four owners tell you nobody loses work because referrals already fill the diary, the packet was wrong and you have been saved months. Do not smooth it over, and do not go looking for a fifth person who will agree with the paperwork. Write the contradiction into the brief as a finding, with the quote that produced it. What comes out of that is a much narrower question than the one you started with: not whether trades miss calls, but which specific owner, in which specific trade, already spends money on this, and what a scoped piece of work would have to be worth to them to be worth a price.