icp-builder

ICP Builder

Run a structured 14-step interview to produce a cold-email-usable ICP. One question per message. Generate ICP v1 (fast) -> ICP v2 (tight after 50-100 sends).

Inputs

• Optional (helpful): company.md from strategy-architect - skip pre-fillable questions when confidence is high/medium.
• Optional: existing icp.md - triggers v2 mode (if v1 exists) or lite-upgrade (if source: strategy-architect-lite).

Workspace path

Default campaign: default. If user invocation includes --campaign <name>, use that. Full rule: see using-outreach-superpower.

All file reads/writes happen in outreach-workspace/<campaign>/.

References

• references/execution-protocol.md - Understand -> Plan -> Confirm -> Execute -> Verify loop. The 14-step interview itself is the Plan + Confirm; Step Final is Execute; Step 15 ratifies the search-ready filter block before downstream skills consume it.

ICP framework (PB-10 lens)

The 14-step interview captures the segment's pain, outcome, and angle. Behind the scenes, every segment must resolve into four filter layers downstream tools (lead-finder, Saleshandy Lead Finder, LinkedIn Sales Nav) can actually filter on:

Layer	What it answers	Captured in steps
Firmographics	Industry, employee range, revenue range, HQ country, funding stage	Steps 5, 10
Technographics	Tools they use today / don't use (competitor signals, stack tells)	Steps 5, 12
Behavioral triggers	Recent events that change buying urgency (hired VP X, raised round, launched product)	Step 7
Intent signals	What they're reading / engaging with that hints at active interest	Step 7, optionally Step 14

Tighten vague replies on these dimensions during the interview. "SaaS" is not firmographics until you ask back-end size + revenue + geo. "They use HubSpot" is technographics. "Just raised Series B" is a trigger. "Downloading cold email content" is intent.

The final Step 15 emits these four layers as a structured YAML block downstream skills consume. Empty layers are allowed.

Operating rules

1. One question per message. Never bundle multiple questions in one turn.
2. Suggestions every question. Always offer 2-3 numbered options + "or reply with your own answer." Think hard about defaults from company.md when available.
3. Visible progress tracker. Format: Step X / 14 (or Step X / N if some are pre-filled and skipped, e.g. Step 3 / 10). Show this on every question.
4. Tighten vague answers. If a reply is vague (e.g. "tech companies", "everyone", "small business"), ask one focused follow-up before moving on.
5. Accept "skip". If the user types "skip", move on and record the field as [SKIPPED] in the final output.
6. Tables in output. ICP Summary, Segment Cards, Targeting Filters, and Angle Kit all render as markdown tables. The user has explicitly asked for tables.

Process

Step 0 - Detect mode

Read the workspace before asking anything:

• If icp.md exists with frontmatter version: v1 AND source: icp-builder-interview -> v2 mode (jump to Step A).
• Else if icp.md exists with source: strategy-architect-lite (and needs_tightening: true) -> lite-upgrade mode: run the full v1 interview but pre-fill heavily from the lite ICP and company.md. On write, replace the lite file (move it to icp.lite.v1.md first) and set source: icp-builder-interview.
• Else -> fresh v1 interview.

Also read company.md if present. Note the confidence flag - it determines pre-fill scope (see the tiered Pre-fill rules table below):

• confidence: high -> 7 steps pre-fillable (1, 2, 5, 6, 8, 9, 13). Show pre-filled answers; user confirms or edits each.
• confidence: medium -> 4 steps pre-fillable (1, 2, 5, 13).
• confidence: low -> re-ask all 14 from scratch; the extracted values aren't trustworthy.

v1 interview (14 steps - preserve exactly)

For each step, output:

• The progress tracker (Step N / 14, or adjusted denominator if pre-filled steps were skipped)
• The question
• 2-3 numbered suggestions (use company.md context when relevant)
• The line: "Reply with a number, or share your own answer."

The 14 questions:

#	Question
1	Tell me your website or what you're selling (1 line)
2	Who is it for today? (current paying users, if any)
3	What problem do they pay you to solve? (in their words)
4	What outcome do they want? (measurable)
5	Your best 10 customers: what do they have in common? (industry/size/geo)
6	Who is a bad fit? (who churns / complains / never buys)
7	What's the "why now"? (trigger events)
8	Who feels the pain daily? (job titles)
9	Who approves budget? (job titles)
10	Typical price point / contract size? (range is fine)
11	Sales motion: self-serve, demo, or both?
12	Key competitors or alternatives they use today?
13	Proof: what can you credibly claim? (results, logos, numbers)
14	Where will you source leads? (Saleshandy/LinkedIn/CSV/etc.)

Pre-fill rules (based on company.md confidence flag):

Confidence	Pre-fill these steps	Behavior
high	1, 2, 5, 6, 8, 9, 13 (7 of 14)	Show pre-filled values; user confirms or edits each
medium	1, 2, 5, 13 (4 of 14)	Show pre-filled values; user confirms or edits each
low (or no company.md)	none	Ask all 14 from scratch

Step-to-source mapping:

• Step 1 (website/product) <- company.md value prop
• Step 2 (current users) <- company.md ICP segments
• Step 5 (best customers commonalities) <- company.md proof / case studies
• Step 6 (bad fit) <- company.md constraints / disqualifiers (high-conf only - explicit data needed, low-conf often missing)
• Step 8 (job titles, pain) <- company.md primary personas (high-conf only)
• Step 9 (job titles, budget) <- company.md primary personas (high-conf only - guess if not explicit)
• Step 13 (proof) <- company.md proof points

For pre-filled steps, render as: "From company.md: . Confirm with y, or share an edit."

Adjust the progress tracker accordingly: Step X / N where N = 14 minus skipped count.

Vague-answer follow-up triggers:

• Single-word industry replies ("SaaS", "tech", "agencies")
• Replies that don't include any number, role, or specific descriptor
• Replies under 5 words on Steps 3, 4, 5, 7, 8, 9
• Step 5 (firmographics): if reply lacks size + revenue + geo, ask back. "Got the industry. What employee range, what revenue band, and which geos?"
• Step 7 (why now / triggers): if reply is generic ("they need it"), probe for a real event - "Trigger events look like: just hired a VP X, raised funding in last 90 days, launched a product, did a layoff, switched a competitor. Any of those fit your best customers?"
• Step 12 (competitors / tech stack): if user names a competitor by category not product, probe - "Got it - which specific tools? HubSpot, Salesforce, Outreach.io? That's the technographic signal we'll filter on."

When triggered, ask one tightening question, then continue.

Step A - v2 mode

Turn 1 - Acknowledge: "Generating v2 (tightening). Send your reply data from the last 50-100 sends when ready - which segments responded, which didn't, what objections came up, any new patterns. I'll wait for your message."

Turn 2 (after user pastes): parse the reply data and run the 5-step tightening interview:

1. Which segment had the highest reply rate?
2. Which segment(s) should we drop?
3. Best new angle that emerged from replies?
4. New disqualifiers learned?
5. New trigger events you noticed?

Update icp.md to version: v2, archive old version as icp.v1.md.

Step Z - Final deliverables

After the last interview question, output all five sections in chat AND write them into icp.md:

A) ICP Summary

Field	Value
Who it's for	...
Who it's not for	...
Primary pain	...
Primary outcome	...
Why now (trigger)	...
Buying roles	...
Proof points	...

B) Segment Cards (3) - one table per segment. Keep segments few (1 primary, 2 secondary).

Field	Value
Segment name	...
Firmographics	...
Technographics	...
Common pain	...
Trigger events	...
Messaging angle	...
Objections + counters	...
Disqualifiers	...
Example titles to target	...

C) Targeting Filters (Saleshandy / LinkedIn ready)

Filter	Value
Industry	...
Company size	...
Geography	...
Keywords	...
Job titles	...
Tech stack signals	...
Funding/intent signals	...

D) Cold Email Angle Kit (5 angles) - one table per angle.

Field	Value
Angle name	...
Subject lines (3)	1. ... / 2. ... / 3. ...
Opening patterns (2)	1. ... / 2. ...
CTA options (2)	1. ... / 2. ...

E) Disqualifiers - bulleted list of who NOT to target (industries, sizes, signals, behaviors).

Every segment must pass the test: "Can I find these people with filters in Saleshandy/LinkedIn?" If not, tighten before writing.

Step 15 - Search-ready filters

After Step Z confirms the ICP body, draft a structured YAML filter block for the primary segment. This is what lead-finder reads to formulate Lead Finder / Sales Nav queries.

Walk the user through it section by section. Confirm each section before moving to the next. Empty sub-keys are allowed (e.g. technographics: {} if no signals exist).

Turn 1 - Firmographics: show the drafted block from the interview answers (Steps 5, 10):

firmographics:
  industries: [SaaS, Fintech]            # from Step 5
  employee_range: [51, 500]              # ask if not captured: "Tightest employee range you'd accept?"
  revenue_range_usd: [1000000, 50000000] # from Step 10 if numeric; else ask
  hq_countries: [US, UK, DE]             # ask if missing: "Which countries / regions?"
  funding_stages: [Series A, Series B]   # optional - only if relevant

Ask: "This is the firmographics block I'd hand to lead-finder. Confirm, edit any line, or say skip to leave empty." Apply edits, re-display, loop until yes.

Turn 2 - Technographics: from Step 12 (competitors / current stack):

technographics:
  uses: [HubSpot, Salesforce]    # tools that signal good fit (competitor switching opportunity)
  not_uses: [Outreach.io]        # tools that disqualify (already solved problem with them)

If user has no technographic signals: technographics: {}. Confirm, then proceed.

Turn 3 - Behavioral triggers: from Step 7. List as snake_case event identifiers:

behavioral_triggers:
  - hired_VP_Sales_last_90_days
  - raised_funding_last_180_days
  - launched_new_product_last_60_days

Format: <event>_<window> so they're filterable. Empty list allowed.

Turn 4 - Intent signals: content / activity hints that suggest active interest. From Steps 7 and 14:

intent_signals:
  - downloaded_cold_email_content
  - active_on_LinkedIn_outbound_topics
  - hiring_for_SDR_or_BDR_role

Empty list allowed if user has no signals.

Turn 5 - Final block confirmation: show the assembled YAML in a code block:

## Search-ready filters

```yaml
firmographics:
  industries: [SaaS, Fintech]
  employee_range: [51, 500]
  revenue_range_usd: [1000000, 50000000]
  hq_countries: [US, UK, DE]
  funding_stages: [Series A, Series B]

technographics:
  uses: [HubSpot, Salesforce]
  not_uses: [Outreach.io]

behavioral_triggers:
  - hired_VP_Sales_last_90_days
  - raised_funding_last_180_days

intent_signals:
  - downloaded_cold_email_content
  - active_on_LinkedIn_outbound_topics
```

Ask: "Confirm? Type yes to write to icp.md, or paste line edits." Loop until confirmation. This block goes at the end of icp.md body, after Disqualifiers.

Step Final - Write file

Write outreach-workspace/<campaign>/icp.md. Frontmatter:

---
version: v1 | v2
generated_at: YYYY-MM-DD
segments: 3
disqualifiers: <count>
source: icp-builder-interview
primary_segment: "<name of segment 1 - the one you'd run a campaign against first>"
buying_motion: self-serve | demo | both
campaign: <name>
---

The primary_segment and buying_motion fields let downstream skills (email-auditor, email-sequence-generator) introspect via frontmatter rather than markdown body parsing. Set primary_segment to the name of the first segment card (the one you'd run a campaign against first). Set buying_motion from the user's Step 11 answer. Set campaign to the campaign name in use.

Body order: ICP Summary -> Segment Cards -> Targeting Filters -> Angle Kit -> Disqualifiers -> ## Search-ready filters (YAML block from Step 15).

Versioning rules:

• v1 fresh write -> just create icp.md.
• v2 -> first move existing icp.md to icp.v1.md, then write the new v2 file.
• Lite-upgrade -> move existing lite icp.md to icp.lite.v1.md first, then write the new full v1 file with source: icp-builder-interview.

End the run with: "ICP v1 saved with structured filter block. Next: run lead-finder to turn these filters into a verified prospect list. (Re-run me after 50-100 sends to generate ICP v2.)"

TodoWrite usage

For the interview, create a TodoWrite list at start with one todo per step that will actually be asked (skipped steps from pre-fill rules don't get a todo) PLUS one todo for "Step 15 - confirm search-ready filter block (4 turns)". Mark in_progress when asking, completed when answered.

Vague-answer follow-ups don't add a new todo. The original step stays in_progress until a non-vague answer arrives.

v2 mode TodoWrite list (7 items):

1. Capture reply data
2. Step 1 - highest reply rate segment
3. Step 2 - drop segments
4. Step 3 - best new angle
5. Step 4 - new disqualifiers
6. Step 5 - new trigger events
7. Generate v2 deliverables and write icp.md

This lets the skill resume mid-interview if context switches.

Abandon and resume

The 14-step interview is conversational. If a user stops mid-interview and returns later, behavior depends on session state:

• Same session, context still in memory: TodoWrite tracks step status; resume from the last in_progress todo.
• New session (no in-memory state): the partial interview is lost. The user restarts from Step 1, but pre-fill rules from company.md shorten the path. No icp.md is written until Step Z is reached.

To improve resume support in v0.2: write a icp.draft.md after Step 7 (halfway through), with status: incomplete frontmatter. On re-invocation, detect the draft and offer to resume or restart.

---

Next: run lead-finder to turn the search-ready filters into a verified prospect list. Re-run me after 50-100 sends to generate ICP v2.