Close-up of hands on a laptop browsing an e-commerce site in a modern office.

How I Build ICPs with Claude (And Why Most Teams Get Them Wrong)

Most ICPs are fictional. A few adjectives, a job title, some assumed pain points. Written once in a strategy offsite. Never updated.

The result: your sales team targets the wrong accounts, your messaging stays generic, your win rate doesn’t move. Not because the product is wrong. Because the ICP was never grounded in anything real.

Here’s how I use Claude to fix that or in other words: How to Build an ICP with AI.


Why Most ICPs Fail Before They’re Used

A typical ICP document looks like this:

  • Industry: SaaS, 100–500 employees
  • Title: VP of Sales or CRO
  • Pain: Struggling with pipeline visibility
  • Budget: $50k–$200k

That’s a market segment, not an ICP. It tells you who could buy. It says nothing about who actually buys, stays, and expands.

In How to Build a GTM Strategy Step by Step, I covered what a useful ICP actually needs to capture: the trigger that makes a customer need a solution now, the cost of inaction if they don’t act, the decision dynamics inside their organisation, and the alternatives they’re currently using. Most ICP exercises stop at firmographics and never get to any of those four.

That’s the gap this workflow closes.


The Four Elements That Make an ICP Actionable

Before touching Claude, you need to know what you’re building toward. An ICP is only useful if it can answer these four questions (full context in the GTM strategy framework):

1. The trigger. What is happening in their world that makes them need a solution right now? In B2B, it’s usually an event: a new hire, a failed launch, a missed quarter, a contract renewal. In FMCG, it’s a consumption moment. A specific occasion or job-to-be-done that drives the purchase. Vague pain is hard to sell against. A trigger is specific enough to prospect against.

2. The cost of inaction. What does it cost them if the problem stays unsolved? Time, money, reputation, or competitive position. Quantify it where possible. If you can’t articulate this clearly, you don’t have a sales motion yet.

3. The decision dynamics. Who uses the product, who pays, who can block it? In B2B, these are rarely the same person. In FMCG or consumer, the buyer and the end user are sometimes different too. Gifting, foodservice, and family purchase decisions all involve multiple stakeholders with different priorities.

4. The alternatives. What are they doing today instead? This tells you what you’re actually competing against. It usually reveals the real objection you need to overcome before anything else.

The narrower your answers, the easier everything that follows becomes. Narrow feels risky. In practice, it’s what makes a GTM plan executable.


My Claude Workflow for ICP Building

This takes about 45 minutes the first time. Quarterly refreshes run in under 20.

Before You Start: Gather Your Data

Claude can only synthesize what you give it. Pull together:

  • 10–20 closed-won deals from the last 12 months
  • 5–10 churned or lost deals
  • CRM fields: industry, company size, deal cycle length, ACV
  • Any customer interview notes or call recordings (summarised is fine)

If your CRM data is messy (missing fields, inconsistent tagging, no close reasons logged), the output will reflect that. Use Prompt 3 below to explicitly flag weak data spots. That gap analysis alone is worth passing to RevOps.

Prompt 1: Pattern Extraction

Start by finding what your best customers have in common.

Prompt
You are a GTM strategist. I’m going to give you data on our closed-won and closed-lost deals. Your job is to identify patterns that distinguish our best-fit customers from poor-fit ones.

Analyze the following:
– What firmographic characteristics appear most often in won deals?
– What buying triggers or events appear in won deals but not in lost deals?
– What objections appeared in lost deals that did not appear in won deals?
– Who was the internal champion in won deals (role, seniority, department)?

Here is the data: [paste your deal data]

Format your output as a structured table: Pattern | Frequency | Strength of Signal (High / Medium / Low)
↑ Copy and paste directly into Claude

Prompt 2: ICP Draft

Feed the pattern output directly into this prompt.

Prompt
Based on the patterns identified above, draft an Ideal Customer Profile.

Include:
– Firmographics: industry, company size, revenue range, geography
– Trigger: what specific event or situation makes them need a solution now?
– Cost of inaction: what does it cost them if the problem stays unsolved?
– Decision dynamics: who uses the product, who pays, who can block the purchase?
– Alternatives: what are they doing today instead of buying from us?
– Champion profile: who internally drives the evaluation?
– Exclusion criteria: who should we NOT pursue, and why?

Be specific. Avoid vague language like “growing company” or “innovative team”. Name the actual signals.
↑ Copy and paste directly into Claude

This structure maps directly to the four elements above. If Claude returns vague answers on trigger or alternatives, the source data is likely thin in those areas. That’s a signal to dig deeper, not a reason to accept the vague answer.

Prompt 3: Pressure-Test It

The ICP draft is only as good as the data behind it. This prompt finds the gaps before your sales team does.

Prompt
Review the ICP above and challenge it.

Identify:
– Which criteria are based on assumptions rather than evidence from the data?
– Which criteria are too broad to be actionable (e.g. could describe 50% of the market)?
– What signals might be missing because we didn’t collect data on them?
– Where does this ICP risk excluding good-fit customers or including bad-fit ones?

For each issue, suggest what additional data or validation would resolve it.
↑ Copy and paste directly into Claude

Most teams skip this step. The output is worth sharing with your sales leader before you build anything around the ICP.

Prompt 4: Scoring Model

An ICP that lives in a document doesn’t change behaviour. This prompt converts it into something a rep can use in five minutes per account.

Prompt
Convert the ICP above into a lead qualification scoring model.

Format:
– List 8–10 criteria that can be assessed from public data or a short discovery call
– Assign each a point value (total = 100)
– Define the threshold for a “qualified” account
– Flag any criteria that should be automatic disqualifiers regardless of total score

Output as a table.
↑ Copy and paste directly into Claude

The scoring model connects your ICP to actual pipeline decisions. It also makes it possible to audit whether your team is following the ICP or quietly ignoring it.


What You Have After Four Prompts

  • A documented ICP with evidence-backed criteria across all four elements
  • A scoring model reps can apply immediately
  • A gap analysis for RevOps on what data to start capturing
  • Explicit exclusion criteria (often the most valuable output)

Two pages. No narrative filler.


When to Refresh

Most teams treat ICP as a one-time exercise. Quarterly is the right cadence. Refresh whenever:

  • Win rate drops more than 10 percentage points
  • You enter a new segment or launch a new product line
  • Churn patterns shift noticeably
  • Deal cycles start getting consistently longer

Run the same four prompts with updated deal data. If you documented the previous version well, the differences between rounds are often the most useful output. They show you how your ideal customer is evolving.


Quick Win: Start Here

Not ready to run the full workflow? Do this first.

Paste 5 recent closed-won deals into Claude (company name, industry, size, deal value, close time) and use this prompt:

Prompt
I’m building an ICP for my GTM team. Here are 5 recently closed deals:

[paste deals]

Based on this limited data, what are the top 3 patterns that seem to characterize our best customers? For each pattern, tell me: is this a trigger, a firmographic, a decision dynamic, or a cost-of-inaction signal? What additional information would help validate it?
↑ Copy and paste directly into Claude

You’ll have a working hypothesis in under three minutes. Good enough to pressure-test in your next pipeline review.


FAQ

What data do I need to start?

Ten closed-won deals from the past 12 months is enough to find initial patterns. More data improves signal quality, but don’t let the perfect dataset stop you from starting.

We’re early-stage with few customers. Does this still work?

Yes, but be explicit about the limitation. Use Prompt 3 to flag which criteria are hypotheses rather than evidence. Treat the output as a testable assumption, not a finished ICP.

How is this different from the ICP section in the GTM strategy guide?

The GTM strategy framework covers what an ICP needs to contain and why. This article is the workflow for building one with Claude, prompt by prompt, using real deal data.

How is an ICP different from a buyer persona?

A buyer persona describes who a person is: demographics, motivations, communication style. An ICP describes which companies to sell to and under what conditions. Both are useful. The ICP comes first.

Do I need a paid Claude plan?

The free tier handles shorter prompts. For multi-step workflows with larger datasets, Claude Pro manages longer context more reliably.


The four prompts above are part of the AI GTM Operator Prompt Pack, a bundle of copy-paste ready prompts for GTM, Sales, and RevOps workflows. Join the newsletter to get early access when it launches.


Related:

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *