Cube Icp Definition

Defining Your ICP

Execution principle

This skill runs silently and completes in a single response. When invoked, Claude's immediate next action is a tool call — not a chat message. No opening statement, no progress narration, no interim findings, no mid-run bailouts, no "want me to continue?" prompts. The user sees tool calls rendered by the UI; Claude's next chat output is the finished deliverable.

If the full analysis can't fit in one turn, narrow the scope silently — but always complete a usable deliverable in one response. Never dump partial findings as an interim summary and ask to continue.

The document is the deliverable. Everything else is noise.

What this skill produces

A cross-functional ICP document delivered two ways — inline in chat so the user can read it immediately, and as a downloadable markdown file they can save to Notion, Google Docs, a shared drive, or wherever their team works.

The document has five sections:

1. ICP snapshot — who the company sells to, in one page. For leadership, marketing, and anyone who needs the high-level answer.
2. Prospect list criteria — primary / secondary / exclude filters, triggers, disqualifiers, and seed accounts. For sales, SDRs, paid ads.
3. Buyer personas — three archetypes with day-to-day reality, what they care about, and where to find them. For sales, marketing, content, paid ads.
4. In-market language bank — verbatim customer lines that indicate an account is ready to buy. For sales qualification, SDR prospecting, marketing keyword research.
5. Competitive landscape — who's in the account and how customers talk about them. For positioning, qualification, paid ads exclusions.

One document, read differently by different teams. Section headers use the human labels — never label sections as "Artifact 1, Artifact 2," etc.

When to use

• Before launching a new outbound motion
• After reaching product-market fit but before scaling
• When entering a new segment or vertical
• When the customer base has evolved and old assumptions are stale
• As part of quarterly ICP reviews

When NOT to use

• Fewer than 50 conversations in Format — not enough signal yet
• Defining ICP from scratch with no existing customers — needs discovery interviews first
• Writing cold email copy or ad creative — downstream skills handle those

Setup

If Format MCP isn't connected yet:

1. Settings → Connectors → Add custom connector
2. URL: https://useformat.ai/api/mcp
3. Authenticate with your Format account

The run — how the skill executes

Tight sequence. Target: ~12 tool calls total. Broad topic-first queries, not over-filtered semantic queries.

Step 1: Orient (1 call)

describe_org()   → the workspace in one call

It returns the org that answered plus this workspace's topics (each with the standing question it asks and its insightCount), the CRM attributes available to filter on and the operators each accepts, the connected sources, the coverage date span, and processing.hasGroups — whether Format has gathered these conversations into insight groups yet. If the connection can reach several workspaces, list_organizations() names them.

Topic structures vary across Format orgs. Map the available topics to the roles below and proceed. Do not surface this mapping to the user. topicNames takes those names case-insensitively; an unknown one is refused with the valid list rather than returning nothing.

Topic role mapping:

If the workspace has the topic, use it. If not, fall back to semantic queries against all topics.

Step 2: Build Best and Worst cohorts (2 calls)

Best cohort — positive-signal topic, no narrow sentiment filter:

search_insights(
  topicNames: [best-cohort topic],
  limit: 75
)

Worst cohort — worst-signal topic:

search_insights(
  topicNames: [worst-cohort topic],
  limit: 75
)

Each row carries the company that said it (with source: 'linked' when Format knows the customer and 'inferred' when the name was only read out of the conversation — the id is null in exactly that case). Extract unique company names from each cohort and dedupe companies appearing in both.

Rows also carry isAiRejected — Format's quality filter's verdict on the extraction. Read it as a caution, not a gate: don't rest a claim on flagged rows alone, but never silently drop a flagged quote whose words are clear — the filter judges extraction quality, not relevance.

Step 3: Resolve companies, firmographics and roles (2 calls)

list_companies(hasInsights: true, limit: 200)
  → match Best/Worst cohort names to company records: IDs (needed for Step 4
    filters) + the attributes array (CRM-mapped fields like industry, plan,
    company size — whatever this org has mapped). `totalCount` says how many
    matched in all, so you know whether one page covered the register.

list_persons(companyIds: [Best cohort company IDs], hasInsights: true, limit: 200)
  → who actually shows up in the conversations; roles from titles/attributes

Two calls. Enough to see dominant industry and dominant buyer role. If the org has no CRM attributes mapped (empty attributes arrays), don't fabricate firmographics — infer what you can from the conversation content pulled in Step 4 and say so in the document.

Step 4: Core content extraction (6–7 calls)

Pull the substance for ICP snapshot, personas, in-market signals, competitive landscape with broad queries against the Best cohort. Do not run separate framework passes.

Why customers chose them (feeds ICP snapshot + competitive landscape):

search_insights(
  topicNames: [competitive/in-market topic],
  companyIds: [Best cohort company IDs],
  limit: 40
)

Pains and what's broken (feeds ICP snapshot, personas, in-market signals):

search_insights(
  topicNames: [product gaps topic],
  companyIds: [Best cohort company IDs],
  limit: 40
)

Value realized (feeds ICP snapshot, personas):

search_insights(
  topicNames: [best-cohort topic],
  companyIds: [Best cohort company IDs],
  limit: 40
)

Onboarding / use case patterns (feeds personas):

search_insights(
  topicNames: [onboarding topic, if available],
  companyIds: [Best cohort company IDs],
  limit: 30
)

Disqualifier content (feeds the Not-Fit account list): reuse the Step 2 worst-cohort pull — it is the same query; don't re-run it.

In-market language sweep (feeds the language bank — always run this one):

search_insights(
  semanticQuery: "why we started looking, what pushed us to evaluate, the problem that made us reach out",
  limit: 30
)

Corpus-wide competitor sweep (feeds the competitive landscape — always run this one):

search_insights(
  keywordSearch: [known competitor and alternative-tool names],
  limit: 40
)

Competitor names surface everywhere — pricing objections, onboarding calls, offhand comparisons — not only under the competitive topic. One unscoped keyword sweep across the whole corpus alongside the topic-scoped pull keeps the landscape built from every mention, not just the ones filed where you thought to look.

The competitive landscape, sized (1 call, when processing.hasGroups is true):

search_insight_groups(
  topicNames: [competitive/in-market topic],
  limit: 20
)

Where Format has gathered these conversations into insight groups, each group is a theme across customers with a customerCount — how many distinct customers raised it. That is the honest basis for the "where it shows up" column: rank alternatives by it, and never add the column up (nested themes count the same customer more than once). search_insights({ supportingGroupId: "<id>" }) returns the words underneath any theme you want to evidence.

Every insight sits under exactly one topic, so a topic-scoped pull only ever returns the topics you thought to name — and pre-purchase language is often filed under one you didn't. One unscoped semantic sweep alongside the topic pulls keeps the language bank from reading only what you guessed at.

6–7 calls. Combined with orientation, cohort-building and company resolution: 11–12 total — pagination on large registers adds more; spend the extra calls rather than truncating the cohort.

Step 5: Synthesize

All five sections built from the pool of extracted insights above. No additional tool calls needed.

The document header states the data window. Right under the title, one line: how many conversations/insights the analysis drew on AND the period they span (earliest → latest timestamp actually read, e.g. "built from ~2,800 insights spanning Dec 2025 – Jun 2026"). A count without a period is the first follow-up question every reader asks — answer it up front.

Adaptation rules

Small cohort (<15 Best cohort companies). Deliver what's supportable. Note — in the "why this ICP, not another" section or the closing calibration note — that confidence would grow with more data.

Different topic names. Map silently via the topic role table above.

Heavy prospect skew (mostly pre-sales conversations). Treat "high-intent prospects who chose us" as the Best cohort. The analysis is about who fits, not how long they've been customers.

Heavy customer skew (mostly existing-customer conversations). The Best cohort degrades into "whoever happened to be on a call recently" unless fit is separated from presence. Recency of conversations is not evidence of fit: pick the Best cohort from retention and expansion evidence — expansion signals and customer-love content, and the CRM attributes where mapped (ARR, renewal outcome, churn score) — rather than from appearing in the window. Say in the calibration note that the corpus skews to existing customers and how the cohort was chosen because of it.

Short coverage window. Outcome- and trend-shaped claims — displaceability ratings, win/loss patterns, geographic or segment trends — need history the window may not hold. When describe_org's coverage span is weeks rather than quarters, don't assert them: state what customers say and under what conditions they'd move, mark anything that depends on pre-window evidence as outside coverage, and name the gated claims in the calibration note.

Different verticals (HR-tech, dev-tools, vertical SaaS, etc). Trigger language varies. Use customer's own words from the data, not generic B2B phrases.

The deliverable — five sections

Two outputs every run:

1. Inline in chat — the full document rendered so the user can read it immediately.
2. A markdown file — the same document saved as [company-slug]-icp.md, delivered however your environment shares files (a download link on claude.ai, a file written to the working directory in Claude Code or Cursor). If your environment can't produce files, skip this and deliver inline only.

The markdown file is important. A GTM source-of-truth document needs to live somewhere — Notion, Coda, Google Docs, a shared drive — not buried in a chat history. Markdown pastes cleanly into every modern tool and preserves the tables.

How to produce the file

After completing the analysis:

1. Save the full document as [company-slug]-icp.md using your environment's file mechanism (on claude.ai: write to the outputs directory and present the file; in Claude Code/Cursor: write it to the working directory).
2. Use the same content for the inline chat response — do not abbreviate either version.

The company slug should be lowercase with hyphens — "Northwind" becomes northwind, "Acme Corp" becomes acme-corp.

Document structure

Both the inline version and the file use this structure:

# [Company] ICP

*This document is the source of truth on who [Company] sells to. Sales pulls the prospect criteria and qualification framework. Marketing grabs the personas and in-market language. Paid ads grabs the filters and exclusions. Product grabs what buyers want and can't get elsewhere. Leadership grabs the snapshot and competitive landscape.*

---

## 1. ICP snapshot
[content]

---

## 2. Prospect list criteria
[content]

---

## 3. Personas
[content]

---

## 4. In-market language bank
[content]

---

## 5. Competitive landscape
[content]

Numbered section headers (1. 2. 3. 4. 5.) are fine and help scanning. Section names use the human labels: ICP snapshot, Prospect list criteria, Personas, In-market language bank, Competitive landscape.

ICP snapshot

Framing subtitle, summary table, narrative paragraph, and a "why this ICP" note. No TAM definition at the top — that's a separate concept from this skill's scope, and cramming it in at the top adds weight without helping.

Open with an italicised one-liner subtitle:

*This document is the source of truth on who [Company] sells to. Sales pulls the prospect criteria and qualification framework. Marketing grabs the personas and in-market language. Paid ads grabs the filters and exclusions. Product grabs what buyers want and can't get elsewhere. Leadership grabs the snapshot and competitive landscape.*

Summary table — 8 rows, specific not generic:

The Emotional state row is what separates this from a generic ICP doc. Pull real customer phrases into it.

Every row in this document's tables is fill-if-supported: where neither the conversations nor the CRM attributes genuinely support a dimension, write "not supported by the data" rather than a plausible value — an honest gap keeps the rest of the table credible.

Narrative paragraph (5–7 sentences, prose only, no bullets):

Cover, in order:

1. Who they are as a company and what triggers the purchase
2. Their current setup and what's broken about it
3. Their existing tech stack and what integration means to them
4. What alternatives they've looked at and why those fell short
5. The wedge — the specific thing the company does that no alternative does
6. Where the ceiling is (above which the deal breaks) and where the floor is (below which the pain isn't acute enough)

No framework jargon. No bullet points inside the paragraph. This reads like a strategist wrote it.

Why this ICP, not another (2–3 sentences):

Name the overlap, not just exclusions. Example structure:

• "Acme wins when the buyer has pain on both sides — dispatch AND invoicing. Pure-dispatch buyers (large fleets) won't switch for scheduling alone. Pure-invoicing buyers (under 10 technicians) don't value the routing and balk at cost. The ICP is the overlap."

This framing is stronger than "we exclude X, Y, Z" because it explains why the ICP is narrow, not just that it is.

Target account criteria

Four components: a single Primary/Secondary/Exclude criteria table, a short disqualifier list with named examples, a seed accounts paragraph, and in-market trigger signals.

Primary / Secondary / Exclude table:

Disqualifiers (hard stops) — as a short list with named-account patterns:

Not a separate table — a prose list of 4–6 bullets, each naming a real example from the data so sales recognises the pattern:

• [Disqualifier 1] ([Named Account] pattern)
• [Disqualifier 2] ([Named Account] pattern)
• [Disqualifier 3] ([Named Account] pattern)

This format makes the list usable — sales hears "oh, this feels like a Sons UK pattern" and knows to disqualify.

Seed accounts — a single comma-separated paragraph, not a table:

Use these to generate lookalikes: **[Company 1], [Company 2], [Company 3], [Company 4], [Company 5]...**

15–30 companies, bolded, comma-separated, one paragraph. This format pastes cleanly into Apollo/Sales Nav/Clay as a lookalike seed list. A table is harder to lift.

In-market trigger signals — short bulleted list:

For SDRs and paid ads retargeting. Observable patterns:

• New [role] hire announcement on LinkedIn
• "Hiring in [region]" posts from [primary-region] HQ
• [Partner/integration] customer base (warm inbound signal)
• Job postings for [specific title pattern]
• Announcements of [trigger event — acquisition / funding / expansion]
• Company listed as customer of [displaceable incumbents]

Buyer personas

Open with one honest framing sentence about the buying committee structure before the persona table. Example:

"[Company]'s buying committee is small: one champion, sometimes one finance co-signer, sometimes one IT/ops person. Everything hinges on Persona 1."

This signals the hierarchy honestly. Three personas of equal weight reads as a lie.

Persona table — three personas maximum, rich rows:

The row that makes this artifact usable for marketing, not just sales:

• What's broken (their words) — verbatim customer lines with slashes between them give marketers drop-in headline material

The Example people row matters too. Pulling real Name (Company) pairs from the best cohort makes the persona concrete. If sales or marketing wants to validate the persona, they can go look at those people's LinkedIn profiles.

In-market language bank

Rename the section "In-market language bank" — that framing reads better than "signals" to a GTM audience.

Opening line:

"Verbatim or near-verbatim phrases from best-fit [Company] buyers. If a prospect says any of these, they're already qualifying themselves."

Table — 10–14 rows, slash-grouped phrases:

Three things that make this format work:

1. Slash-grouped variations in a single row — instead of one row per quote, group 2–4 variations of the same pattern. This compresses the table and shows the pattern is widespread.
2. "Who uses it" column names real people from the data — not just the persona archetype. "Ops champion (Dana at [Account])" is more useful than "Sales qualifies." Use the real names from the workspace in the document you produce; this skill's own examples stay generic on purpose.
3. "What it tells you" is specific and actionable — "Displaceable incumbent relationship — Acme's wedge" is better than "Ready to buy."

Competitive landscape

Table — 6–10 rows, covering both named competitors and non-obvious alternatives (status quo, adjacent categories):

Three things to include:

1. Rows for NAMED competitors — the tools prospects bring up by name in evaluations
2. Rows for LEGACY/ENTERPRISE incumbents — the consultancies or legacy vendors already holding the account; often the real competition, not the named SaaS alternatives
3. Rows for NON-OBVIOUS alternatives:
   • Adjacent category tools that get confused with the product (e.g. "route-planning apps — logistics tools, not field-service management")
   • Status quo as an explicit row: "Spreadsheets + email + separate provider portals — the actual incumbent in roughly half of deals"
   • Platform-native alternatives that might be perceived as competitors but aren't

Name the real ones from the workspace's own conversations. The categories are the durable part; the names change per company, which is why they aren't listed here.

The Displaceable column should be a paragraph per row, not just "High/Medium/Low." Include:

• When it's displaceable (specific conditions)
• When it's not (specific conditions)
• [Company]'s angle against it (the specific wedge)

Example of a rich Displaceable entry:

"High for mid-market (50–500 employees), Low for large enterprise. Best wedge: 'we ARE the dispatch layer — and we invoice from the same job record.'"

This is more useful than a single word rating because it tells sales exactly when to fight and when to walk away.

The rating itself is gated on coverage: displaceability is an outcome claim, and when describe_org's coverage span is weeks rather than quarters, replace High/Medium/Low with the conditions customers describe — no rating word — and say so in the calibration note (see the Short coverage window adaptation rule).

Close with next-steps offer

After the full document is displayed inline, add the next-steps offer as a single sentence picking 3–4 from this standard menu (do not invent custom options based on findings):

1. Pull a live prospect list in Apollo/Clay/Sales Nav from the ICP criteria
2. Draft persona-specific outbound templates
3. Build an SDR qualification framework from the triggers and disqualifiers
4. Brief marketing on positioning and landing page updates from the personas
5. Run win/loss on specific deals
6. Set up a monthly ICP refresh cadence

One sentence.

A short methodology or calibration note is welcome after the next-steps offer where it helps the reader weigh the document — a line or two on how the cohorts were built, what the data could and couldn't support, and where confidence is thin. Keep it to a few sentences: the document is the deliverable, and the note exists to calibrate it, not to narrate the run.

Still banned at the end of the deliverable (the note calibrates the document; it never describes the assistant's process):

• Rerun offers — "If you'd rather base this on... say the word and I'll rerun"
• Run narration — "I used a hybrid of X signals because..."

Final step: present the file

After the next-steps offer, surface the saved .md file (download link or file path, per your environment). Example:

On claude.ai: write the file to the outputs directory and present it so a download link renders. In Claude Code or Cursor: write northwind-icp.md to the working directory and state the path in one line.

No commentary around the file presentation. The user sees the file, and can save the document to Notion or their shared drive.

The run is complete after the file is presented. Nothing follows it.

Scope boundaries

• Runs silently, completes in one response
• Uses tables throughout — no bulleted prose deliverables
• Methodology stays brief — a short closing calibration note is fine; extended narration of the run is not
• No framework names (JTBD, Four Forces, positioning frameworks) in the output
• No section labels like "Artifact 1" — use human names only
• No cold email copy, subject lines, ad creative, or landing page copy — those are downstream skills

Quality bar before shipping

• Zero chat output between the user's request and the ICP snapshot
• All five sections present (or scope narrowed with a one-line note, not a "continue?" bail-out)
• Every section uses tables as the primary structure
• ICP snapshot stands alone — could be pasted into a positioning doc
• Prospect criteria copy-paste ready for Apollo
• Personas usable by sales, marketing, and paid ads — not just SDRs
• Competitive landscape has Displaceable? ratings (or, under a short coverage window, the conditions customers describe in place of a rating)
• Closing line offers next steps from the standard menu
• Markdown file [company-slug]-icp.md saved and surfaced (download link or file path) as the final action — skipped only if the environment cannot produce files