# One Question, Six Ways: How I Use Claude for Research
*Published: 2026-11-21*
*Tags: ai, agents, for-executives, for-founders*
*Source: https://chrislema.com/six-levels-of-working-with-claude*
---## How I Got There

I've been helping Premom rebuild their website. Premom makes a fertility app and sells ovulation and pregnancy tests. Before we design anything, we need to know two things. What does every site in this space do, so Premom looks like it belongs? And what does Premom do better than everyone else, so we can make sure it stands out?

So I did a competitive analysis. A deep one. Ten websites crawled page by page. Screenshots of every homepage. Tens of thousands of customer reviews from earlier studies we'd already done. Two clear lists at the end.

When the team saw it, they asked a fair question. How did you get there?

They weren't asking about the findings. They were asking about the work. It looked like weeks of research. It wasn't. Most of it was one long working session, and the data collection cost a few dollars.

The honest answer is that I use Claude differently than most people do. It's not that I write magic prompts. It's that I don't treat Claude like a search box. I give it a job, the tools to collect real data, and a team of helpers to read that data. Then I ask it to answer from the data, not from what it remembers.

There are steps between those two ways of working. I count six. And instead of just describing them, I ran the same question through all six so the team could see the difference for themselves.

Here's the question:

> What does Premom.com need to do to compete with other fertility and ovulation-test websites, and where can it win?

Same question every time. Only the instructions around it changed. I ran each level as a fresh session that knew nothing about the others, so the early levels couldn't borrow from the later ones.

## The One Idea Behind All Six Levels

Before I walk through them, here's the point of the whole thing.

At the first few levels, Claude answers from what it knows, or from whatever it happened to see on a few web pages. At the top levels, Claude answers from a pile of real data you collected and saved.

That's the shift. From "what does Claude know?" to "what does our data say?"

The first kind of answer sounds right. The second kind can be checked.

## Level 1: Ask a Question

**What you type:** The question. That's it.

**What happens:** Claude thinks. No web, no files, no tools. It answers from what it learned in training.

**What you get:** A good answer, honestly. It split the market into the right groups. It said Premom needed to earn trust fast, show proof, explain which test to buy, and compete on price per test. It even pulled in some history about the market from memory.

**The limit:** Every point is an idea, not a finding. When I asked it to list what it was guessing about, it was honest. It didn't know what Premom sells today, what things cost, or what the site looks like right now. It guessed.

**Time and cost:** About a minute. Free.

**Use it for:** A first list of ideas before you look at anything.

## Level 2: Send Claude to One Website

**What you type:**

> Go to premom.com. Read the homepage and the main pages it links to: the app page, the shop, the About page, and the success stories. Then tell me: what does Premom.com need to do to compete with other fertility and ovulation-test websites, and where can it win?

**What happens:** Claude opens five pages and reads them.

**What you get:** Now the answer is about the real site. It caught claims that didn't match from one page to the next. It noticed the site made promises it never backed up with proof. It flagged basic shopping details that were missing.

**The limit:** In its own words, every "where it can win" point was "a guess about the competition, not a verified gap." It never looked at a competitor. It also only saw a summary of each page, not the full page.

**Time and cost:** Under two minutes. Free.

**Use it for:** Auditing one page or one site. It's good at "what's wrong with our site?" It can't tell you how you compare.

## Level 3: Sub-Agents Review a List of Sites

**What you type:**

> Here's a list of 9 sites: premom.com and eight competitors. Use a sub-agent for each site. Each one should read the homepage and key pages and report back the main promise, who it's for, the proof it uses, prices, and what stands out. Then compare them all and answer the question.

**What's a sub-agent?** It's a helper copy of Claude. The main Claude starts it up and gives it one job, like "read this one website and report back." Each helper starts with a clean slate, works on its own, and hands back a short report. The main Claude only sees the reports.

**What happens:** Nine helpers start at the same time, one per site. Each one reads about five pages and reports under the same headings. Then a fresh Claude reads the nine reports, builds a comparison table, and counts.

That "same headings" part matters more than anything else at this level. Because every helper reported the same things in the same order, the main Claude could line them up and count. "Seven of eight competitors do this. Premom doesn't."

**What you get:** You can finally see what's normal in the market. A missing trust signal sounds minor on its own. It looks very different when almost every competitor has it. And it found something no single-site read could. The biggest threat wasn't the biggest brand. It was a lookalike competitor selling nearly the same thing.

**The limit:** It only read what companies say about themselves. And its counts depended on how many pages each helper read. Later, a deeper crawl of the same sites changed one of its counts a lot.

**Time and cost:** About three minutes, because the helpers work at the same time. Free.

**Use it for:** A fast side-by-side view when you already know who your competitors are.

## Level 4: A Workflow Picks the Sources

**What you type:**

> I want to know what real customers say, not what companies say about themselves. Pick the places online where women who use fertility apps and ovulation tests leave reviews, comments and testimonials, for Premom and its main competitors. You pick the sites. Send an agent to each one to read the testimonials, comments and reviews and bring back lists and summaries.

Two things changed. First, I stopped picking the sources. Second, I asked for what customers say, not what companies say. That second line changes the evidence completely.

**What's a workflow?** It's a planned set of steps where one step feeds the next. Here there were three: plan, read, combine.

**What happens:**

1. A planner scouts the web. It picks the sources: app store reviews, review sites, parenting forums, roundups that quote real customers. It tests that each one loads. It writes the exact instructions each reader should follow. And it lists what it rejected and why.
2. Four readers go out at the same time and follow those instructions.
3. A fresh Claude combines what they found.

**What you get:** The stuff no company website will tell you. Customers describing a moment of confusion the company had never addressed. Review sites where the brand didn't show up at all. Competitors whose customers keep complaining about the same thing. This is the information a website rebuild needs most.

**The limit:** It read a couple hundred reviews and posts. That's a small sample. And the two places where most of this conversation happens, Reddit and Amazon, were both blocked. The planner said so. Its last note was basically: run this again with a tool that can get through.

That's Level 5.

**Time and cost:** About seven minutes. Free.

**Use it for:** Finding the customer's voice when you don't know where to look. Treat the counts as hints.

## Level 5: Apify Collects the Data, Claude Reads All of It

**What you type:**

> Pick the sites. Use Apify to collect massive amounts of customer data. Then read all of it and write a report that answers our core question.

**What's Apify?** It's a service with thousands of ready-made web robots. Apify calls them actors. You tell an actor what to collect, like "the newest reviews for this app on Google Play," and it puts the results in a table. Claude connects to Apify directly. It picks the actor, checks the price, sets a spending cap, runs it, and pulls the data back.

**What happens:** This level adds two new steps: collect and count.

1. **Pick the sources.** App store reviews for Premom and six competitors, plus Reddit.
2. **Collect.** The Apify robots pull thousands of reviews and posts, each with its star rating and date.
3. **Save.** Everything goes into files, one review per line.
4. **Count with code.** Claude writes short programs to count themes and track ratings month by month. Code doesn't skim, guess or get tired. It treats review number 4,000 the same as review number 1.
5. **Read every line.** Nine reader helpers split the files and read every single review, start to finish. Then a fresh Claude reads the counts and the reader reports, checks a sample of quotes against the raw files, and writes the report.

**What you get:** Answers with numbers and dates you can check. One finding came from code counting dated reviews month by month. That's a pattern no one skimming a few pages would ever see. It also found unhappy customers of other brands naming Premom as the better choice. One said they would have switched if they'd known Premom sold a certain kind of test. Premom does sell it. That customer just didn't know.

**The limit:** Mostly recent reviews, because app stores limit how many they show. The Reddit run hit its time limit early. Keyword counts are rough. And one reader got the dates wrong on two quotes. The final check caught it.

**Time and cost:** About 15 to 20 minutes. About $1.50 in Apify charges.

**Use it for:** Answers you need to defend in a meeting.

## Level 6: Build on What You've Already Collected

This is the real analysis the Premom team saw. I didn't rerun it.

Here's the prompt, almost word for word:

> This is about a new version of the Premom.com website. To build the skills and agents to build a v2, we need data. We need their top competitors listed and their websites analyzed. We need to know the navigation, the copy, the target audiences, the messages, use of imagery and testimonials and more.
> 
> But Premom is different than pure app providers. They also sell products directly from their site. So we also have to look at the top sellers who also sell on their own sites.
> 
> Because we have to have two lists. What we do like everyone else, so that our site reflects it like everyone else. This proves we belong. But it doesn't help you stand out. So the second list is what we do well, beyond everyone else. And make sure those stand out.
> 
> Now, we've done some comparative analysis already on Premom and their product reviews and their Amazon reviews. Those findings are in artifacts that you should be able to find and read.
> 
> Then I want you to pick no more than ten sites and use Apify to pull down their pages, content, designs and screenshots so you can analyze everything.
> 
> After all that, we'll define a Premom Brand Design Guide, a skill to match, a copywriting skill for the website, and skills for anything else we need, and then build out a first pass at v2.
> 
> Make sense? Spend no more than $50 at Apify.

Look at what's in there that isn't in the other prompts:

- **The goal behind the question.** Not "how do we compete?" but "we're rebuilding the site and need to know what to copy and what to feature."
- **The shape of the answer.** Two lists.
- **What makes this client different.** They sell products, so the competitor set has to include sellers, not just apps.
- **What already exists.** Earlier studies, and the instruction to go find them.
- **What to collect, and a budget.** Ten sites. Fifty dollars.
- **What happens next.** The answer feeds a brand guide, writing skills and a site build.

**What happens:** Claude reads the earlier studies first. Then it crawls ten sites deeply, takes screenshots, pulls design details, checks every competitor claim against the crawled text, and builds the two lists.

**What you get:** An answer to the business question, not just the research question. And a plan for the next round of work.

There's one more thing this level shows. The earlier studies it builds on had already surfaced an issue. The team acted on it. Then Level 5, starting from scratch with no access to those studies, found the same thing in its own data. When two separate runs agree, you can trust the finding. And when earlier research has already changed a decision, you know it was worth doing.

**Time and cost:** One long working session, plus the earlier studies. A few dollars in Apify charges.

**Use it for:** Big decisions you'll build on. A site rebuild. A launch. A new product line.

## How the Prompt Changed

Here's what got added at each level:

- **Level 1:** The question.
- **Level 2:** A source, and which pages to read.
- **Level 3:** A list of sources, one helper per source, and a fixed report format.
- **Level 4:** "You pick the sources," and what kind of evidence counts.
- **Level 5:** A collection tool, the instruction to read all of it, and a report that answers one question.
- **Level 6:** The business goal, the shape of the answer, earlier work, a budget, and what happens next.

Three things to notice.

First, the question never changed. Everything around it did.

Second, at Levels 1 through 3, you tell Claude where to look. At Levels 4 through 6, you tell Claude what you need, and it decides where to look. That only works because the planner explains its choices and tells you what it rejected.

Third, the Level 6 prompt reads like something you'd say to a person. It explains why. It ends with "Make sense?" You don't need special words. You need to give the whole picture.

## Why Helpers Change the Work

When you ask one Claude to read nine websites, it has to hold all nine in its working memory at once. Pages are long. The more it reads, the easier it is to lose details from the first few sites.

Sub-agents fix that. Each helper has one job and a clean slate. They work at the same time. And the main Claude works from nine short reports instead of 45 long web pages, which leaves it room to compare and think.

A workflow adds order. The planner's work shapes what the readers do. Without that step, four readers come back with four different kinds of reports.

And at Level 5, every tool does what it's best at. Apify collects. Code counts. Helpers read. One fresh Claude writes the answer.

## Why Apify Changes the Work

Before Apify, Claude reads the web the way you do. One page at a time, and only the pages that load. App stores show a handful of reviews. Some sites block you completely.

Apify changes that. Thousands of reviews instead of a handful. Places you couldn't get into before. Every record complete, with a rating and a date. And every run is saved, so you can see exactly what was collected and when.

It's also cheap. All of Level 5 cost about a dollar and a half.

But Apify doesn't analyze anything. It only collects. The thinking still happens in Claude.

## The Trap at Level 1

Most people stop at Level 1 because the answer sounds right. And it usually does sound right.

But a confident answer from memory and a confident answer from thousands of real reviews look exactly the same on your screen. Only one of them can tell you what changed last month.

## Three Habits That Catch Mistakes

Every level got something wrong. Level 1 guessed. Level 2 only saw summaries. Level 3's counts depended on how many pages it read. Level 4 got blocked. Level 5 had a reader get dates wrong.

Here's how you catch that:

1. **Ask every level to list what it guessed and what it couldn't see.** It will tell you.
2. **Ask for counts and dates, not adjectives.** "Many customers complain" can't be checked. Something like "about 40 of the 250 newest reviews" can.
3. **Make the last step a check.** Have Claude check its quotes and numbers against the raw data before you use them.

## How to Move Up a Level

You don't have to jump to Level 6. Move up one level at a time, and use the level the decision deserves.

**To get to Level 2,** paste a URL, say which pages to read, and end with: "List anything you assumed that wasn't on the page."

**To get to Level 3,** give a list and a fixed report format:

> Use a sub-agent for each of these sites. Each one reads the homepage and up to four key pages, then reports in 350 words under these headings. Then compare them in a table and count how many do each thing.

**To get to Level 4,** describe the evidence, not the sites:

> I want to know what customers say about this, not what companies say. You pick the sources. Plan first: tell me which sources you chose, why, and which you rejected. Then send an agent to each one with clear instructions. Bring back lists with short quotes and dates.

**To get to Level 5,** connect Apify to Claude once, then:

> Pick the sources. Use Apify to collect the reviews and posts. Check each actor's price first and keep the total under this amount. Save everything to files. Count themes and ratings with code. Have agents read every line. Then write a report that answers this question. Check quotes against the raw data before you finish.

**To get to Level 6,** write the prompt like a brief to a smart person. What you're really trying to do. What shape the answer should take. What you already have. What to collect, and the budget. What the answer is for.

Here's my rule of thumb. If you'd be embarrassed to defend the answer in a meeting, you used too low a level.

## What the Team Really Asked

When the Premom team asked how I got there, they weren't really asking about Claude. They were asking whether they could do it too.

They can. The tools are the same tools everyone has. The difference is the habit of moving up. Stop asking Claude what it knows. Start asking it what your data says.
