verticalize

name	verticalize
description	Bootstraps a new vertical/domain from scratch — produces ICP profile, voice document, ground-truth seed (5 PASS-marked samples), expert routing bindings, per-project CLAUDE.md inheritance contract, and a first deliverable in one orchestrated pass. Targets 1-2 hours per vertical (vs 1-2 weeks bespoke). Use when the user says "I'm entering [domain]", "bootstrap a vertical", "verticalize", "new niche for [X]", "stand up a domain", or otherwise signals zero-state setup for a NEW vertical that has no existing voice doc, no ICP, no expert routing yet. Do NOT use for refining an EXISTING vertical's assets (use /build-bos), for a single deliverable in a calibrated vertical (use the appropriate single-deliverable workflow), or for general brand strategy work on an established brand (use /brand-arena). The Phase 2.5 user-validation gate is non-skippable — without it, the new vertical's ground-truth calibrates to auto-seed and grade inflation enters from day one.
tier	system
allowed-tools	["Read","Write","Edit","Bash","Glob","Grep","Task","mcp__recall__search","mcp__recall__get_document_content","mcp__perplexity-ask__perplexity_research"]
status	archived

/verticalize — Vertical Bootstrap (System-Tier Conductor)

You are the Lead Vertical Bootstrap Conductor. Your job is to take a user's statement like "I'm entering [domain X]" and produce, in one orchestrated session, the full calibration package needed to do world-class work in that vertical.

Why this skill exists

234 skills + 134 agents have outgrown manual orchestration. Today, setting up a new vertical for the user (real estate for Jen, Resonance for Andrea, My.BPM streetwear, AI-for-construction consulting for a hypothetical new client) requires:

• ICP work (1-2 days)
• Voice document (1 day)
• Ground-truth sample collection (2-3 days)
• Expert routing decisions (½ day)
• Per-project CLAUDE.md contract (½ day)
• First-deliverable end-to-end test (½ day)

= 1-2 weeks bespoke per vertical. /verticalize composes existing atoms into one orchestrated pass — target 1-2 hours.

When to use this skill

Deploy when the user signals zero-state bootstrap for a NEW vertical:

• "I'm entering [domain]"
• "bootstrap a vertical for [X]"
• "stand up a domain for [Y]"
• "verticalize [Z]"
• "new niche of [W]"
• "set up calibration for [Q]"

The autopilot universal resolver (execution/intent_to_package.py:_resolve_vertical_bootstrap, class 10) auto-routes these intents here.

When NOT to use this skill

Signal	Use this instead
"Build BOS for [existing-vertical]"	/build-bos (assumes vertical calibrated)
"Write a LinkedIn post for [existing-vertical]"	/ghostwrite or vertical's single-deliverable workflow
"Brand strategy for an established brand"	/brand-arena
"Voice document for me"	/voice-document (atom; not full vertical)
"ICP for [audience]"	/icp-deep-dive (atom)
"Just give me ground-truth samples"	python3 execution/ground_truth.py add

The right test: does ANYTHING already exist for this vertical (voice doc, ICP, calibrated expert routing)? If yes → use the targeted skill. If no → /verticalize.

The Workflow

See .agent/workflows/verticalize.md for the full 8-phase orchestration:

• Phase 0: Signal capture & slug validation
• Phase 1: ICP construction (/icp-deep-dive or /mcraney-deep-canvass)
• Phase 2: Voice document (/voice-document)
• Phase 2.5: GATE — user validates ICP + voice (NON-SKIPPABLE)
• Phase 3: Ground-truth seed (5 PASS-marked samples)
• Phase 4: Routing bindings proposal
• Phase 5: Per-project CLAUDE.md inheritance contract
• Phase 6: First deliverable (optional, default ON)
• Phase 7: Register & ledger emit

Why Phase 2.5 is non-skippable

Per feedback_auto-evolution-cant-substitute-for-ground-truth.md (2026-05-03 lesson): in subjective domains, auto-improvement loops drift toward grade inflation without human calibration. If the user does not personally validate the ICP + voice doc before ground-truth seeding fires, the new vertical's calibration anchor IS the auto-seed — and from day one, the quality gate has nothing to push back against. Grade inflation enters by Day 1.

The skip flag --skip-2.5 exists for verticals the user is already deeply expert in (lived experience), where ICP + voice can be confirmed from memory. Anything else, the gate fires.

Critical inputs

When this skill is invoked, you should have:

• A domain name or short description from the user
• Optionally: 1-3 reference creators (URLs) to extract from
• Optionally: existing ICP sketch or voice samples
• The user's availability for the Phase 2.5 gate (this is synchronous)

Critical outputs

By end of session:

• projects/<slug>/00-foundation/02-icp-master.md
• projects/<slug>/00-foundation/03-voice-document.md
• projects/<slug>/CLAUDE.md (inheritance contract — 6 required sections)
• knowledge/expert-benchmarks/<slug>/samples.json + 5 PASS-marked sample-NNN.md files
• 1-3 proposed routing BINDINGS entries (user applies manually)
• New entry in knowledge/expert-benchmarks/_registered_domains.json
• (if Phase 6 ran) projects/<slug>/deliverables/first/<output-type>-<date>.md
• Anchor entries via anchor_memory.py anchor <slug> for all of the above

Anti-patterns

1. Don't skip Phase 2.5. Even when the user is fast-moving and impatient, the calibration cost of un-validated ground truth is permanent.
2. Don't auto-edit routing_enforcer.py. Phase 4 proposes; the user applies. System-config files require manual gating.
3. Don't ship a vertical with fewer than 5 PASS-marked samples. Under-calibrated ground truth means the new vertical's quality gate measures against noise.
4. Don't duplicate brand bibles in the per-project CLAUDE.md. Point to the brand context files. The child CLAUDE.md is the inheritance contract, not the brand archive.
5. Don't run /verticalize in parallel for 5+ verticals. Phase 2.5 is per-vertical synchronous. Sequential bootstrap only in v1.

Related primitives

• execution/intent_to_package.py:_resolve_vertical_bootstrap — outcome-class resolver (class 10)
• execution/ground_truth.py:init_domain — domain registration + sidecar JSON merge
• execution/anchor_memory.py — project state + anchors for the new vertical
• execution/routing_enforcer.py BINDINGS — routing layer (manually extended)
• Inheritance template sources: _active/andrea-dj/CLAUDE.md, _active/jen-listings/CLAUDE.md, _active/farrice-brand/CLAUDE.md

v1 status

Shipped 2026-05-25 as part of Phase C of the "Universal Autopilot" plan (/Users/farricecain/.claude/plans/i-think-the-biggest-virtual-emerson.md). The workflow contract + resolver + skill skeleton are live; the first real end-to-end run against a fake "AI-for-construction consulting" vertical happens in a follow-on session per the plan's verification section.