AI-Powered Partnerships

The Problem

What partnerships work looked like

Before — Manual

✗ 45+ minutes per partner profile: open 6 tabs, copy-paste across HubSpot, Slack, Fireflies, Notion
✗ Co-sell emails: read HubSpot task, find Slack message, look up AM history, compose email, CC the right people
✗ Partner data goes stale: people change jobs, titles change, emails bounce — nobody notices for months
✗ Pipeline review: manually check each deal for meeting status, flag stale ones, ping owners individually
✗ Call notes: listen to recordings, figure out which companies were discussed, update the right HubSpot records
✗ Account mapping: CSV uploads, manual association labeling, conflict resolution — error-prone at scale

After — Automated

✓ Partner Intelligence builds a full 360° Notion page in ~2 minutes from 6 sources in parallel
✓ Co-sell drafts created end-to-end: tasks pulled, Slack content fetched, emails personalized, Gmail drafts ready
✓ Partner cleanup validates every contact against LinkedIn, updates titles, re-maps company associations
✓ Deal notifier classifies every deal daily, sends Slack threads grouped by owner with specific actions
✓ Call notes processed automatically: companies extracted, matched to HubSpot, AM verified, notes written
✓ Account mapper handles CSV parsing, domain resolution, 3-phase association management — zero manual CRM work

System Architecture

How it all connects

Every skill is a .md file in ~/.claude/commands/. Claude Code reads the skill file, plans its approach, then orchestrates API calls across multiple systems. Two HubSpot tokens handle different API scopes. macOS LaunchAgents schedule recurring skills.

Data Flow

Sources

HubSpot CRM

Custom Objects

Fireflies

Slack

LinkedIn/Apify

Web Search

↓ parallel fetch ↓

Processing

Claude Code (Opus)

→

Merge & Classify

→

Synthesize

↓ write ↓

Outputs

Gmail Drafts

Notion Pages

HubSpot Notes

Slack Messages

HubSpot Updates

Scheduling

LaunchAgent Plist

→

Launcher Script

→

claude -p "/skill-name"

Dual token architecture: standard HubSpot token for contacts/deals/tasks + custom object token for Partner Contact (2-40826561) associations

The Skills

13 automations, built to compound

Each skill is purpose-built for a specific partnerships workflow. They share patterns — parallel fetching, graceful failure, human-in-the-loop drafts — but each solves a distinct problem. Grouped by function: Intelligence, Email, CRM, and Infrastructure.

Intelligence & Profiles3 skills

Build and maintain deep knowledge about every partner

Partner Intelligence

Flagship

Builds a complete 360° partner profile in ~2 minutes from 6 data sources — the work that used to take 45+ minutes of tab-switching and copy-paste.

Resolves the partner across HubSpot standard contacts AND the custom Partner Contact object, then fetches everything in parallel: patch companies, co-sell deals, team hierarchy, notes, emails, Fireflies transcripts, Slack mentions, and web signals. Synthesizes it all into a structured Notion page with an action queue, activity log, co-sell intelligence, patch territory map, and team peer links.

Before Open HubSpot, Slack, Fireflies, Notion in separate tabs. Copy data. Format manually. 45+ min per profile.

After One command: /partner-intelligence build partner@aws.com → full Notion page in ~2 min.

1Resolve partner across custom object (primary) + standard contact (secondary)

2Parallel fetch: companies, deals, team, notes, emails, transcripts, Slack, web

3Batch-read details for all entities (single API call per type)

4Merge with field authority rules (custom object wins for 21 fields)

5Classify engagement: Hot (<14d) / Warm (14-60d) / Cold (60d+)

6Write structured Notion page with metadata for incremental updates

// Step 1: Dual-source identity resolution
parallel {
  custom_obj = curl("custom-objects/2-40826561?email={input}")
  std_contact = hubspot.search("contacts", email)
  notion_page = notion.search(email, "collection://27e9...")
}

// Step 2: Parallel data fetch (all in one message block)
parallel {
  companies  = curl("associations/patch-companies")
  deals      = curl("associations/co-sell-deals")
  team       = curl("associations/manager-reportees")
  notes      = hubspot.get("notes", last_10)
  emails     = hubspot.get("emails", last_10)
  transcripts= fireflies.search(partner_name, last_5)
  slack_hits = slack.search(partner_name, last_10)
  web_intel  = web.search(partner_name + company)
}

// Step 4: Field authority — custom object wins
for field in [name, email, title, linkedin, phone,
             city, country, region, segment, lifecycle...]:
  merged[field] = custom_obj[field] ?? std_contact[field]

// Step 5: Engagement classification
days_since = daysBetween(last_activity, today)
engagement = days_since < 14 ? "Hot"
           : days_since < 60 ? "Warm" : "Cold"

// Step 6: Write with incremental metadata
notion.write(page, {
  action_queue, activity_log, cosell_intel,
  patch_territory, team_peer_map, profile,
  metadata: { last_full_build: today, update_count: 0 }
})

HubSpot MCP Custom Object API Fireflies Slack Notion Web Search

Prospect → AM Bridge Intel

Intel

When a prospect responds, instantly builds the context package your AWS AM needs — social proof, engagement history, and a draft message — so the intro feels warm, not cold.

Extracts the prospect signal, finds their company in HubSpot, retrieves the mapped AWS Account Manager (from aws_account_manager_email), builds social proof from similar Clazar customers, fetches the AM's last 3 notes and emails, and synthesizes a ready-to-send reachout plan with multiple angles.

Before Prospect responds → scramble to find AM, look up history, figure out what to say. 20+ min.

After One command → full AM context package with social proof, draft message, engagement status. 2 min.

1Find prospect + company in HubSpot, retrieve aws_account_manager_email

2Build social proof: match Clazar customers by ecosystem (voice AI → meeting intelligence)

3Fetch AM's last 3 notes + 3 emails, classify engagement (Hot/Warm/Cold thresholds)

4Synthesize: prospect signal + AM status + social proof + multi-angle draft message

HubSpot MCP Custom Object API

Partner Team Lookup

Intel

Given any partner contact, instantly maps their full team: manager, all peer reportees, lifecycle stages, account counts, and latest meeting dates.

Traverses HubSpot's custom Partner Contact object to resolve the manager/reportee hierarchy in both directions. If you give it an AM, it finds their SM and all peers. If you give it an SM, it returns all their reportees. Outputs a clean table with lifecycle, deal counts, and meeting freshness.

Before Navigate custom objects manually, click through each association, cross-reference contacts. 15+ min.

After One command → formatted team table with full context. 30 seconds.

1Fetch partner record + partner-to-partner associations (Manager/Reportee labels)

2Resolve direction: AM has Manager, SM has Reportees

3Batch-read all team member details in single API call

4Cross-reference HubSpot contacts for latest_meeting_date

Custom Object API HubSpot MCP Keychain

Email Automation5 skills

Turn CRM tasks and signals into personalized Gmail drafts — humans approve, never robots

AWS Co-Sell Email Drafter

Email

Fully automated end-to-end: pulls HubSpot co-sell tasks, fetches Slack email templates, groups by AWS AM, personalizes with engagement history, and creates Gmail drafts. No approval step between AMs.

Processes Simran's "Co-sell Initiation" tasks: resolves her owner ID, pulls NOT_STARTED tasks, fetches associated deals (filtering out Channel Partnership), parses Slack message links (converting p1773854406648669 to API timestamp format), researches the last 3 months of engagement per AM, composes personalized drafts with the Slack content as the email body, and marks every task COMPLETED.

Before Open each task, find the Slack message, look up AM history, compose email, CC the right people. 10+ min per AM.

After One command processes all pending tasks → Gmail drafts ready, tasks marked complete. ~1 min per AM.

1Pull NOT_STARTED "Co-sell Initiation" tasks assigned to Simran

2Fetch associated deals, filter out Channel Partnership deals

3Parse Slack link from task body → fetch message content via Slack MCP

4Group deals by aws_account_manager_email (multi-deal = single email with bullets)

5Check for active Gmail threads (last 7 days), research personalization (3 months lookback)

6Compose draft with personalization opener + Slack content + booking link → CC partnerships@ + PSM

7Create Gmail draft, mark tasks COMPLETED

// Slack timestamp conversion (critical detail)
raw = "p1773854406648669"
ts  = raw[1:10] + "." + raw[10:]  // → "1773854406.648669"

// Group deals by AM for combined emails
groups = groupBy(deals, d => d.aws_account_manager_email)

for [am_email, am_deals] of groups:
  recent_thread = gmail.search(am_email, last_7_days)
  intel = parallel {
    hubspot.notes(am_email, 3_months)
    fireflies.search(am_email)
    slack.search(am_email)
  }
  draft = compose({
    personalization: intel,
    body: slack_content,  // preserved from Slack message
    booking: "clazar.chilipiper.com/me/caleb-mills/...",
    cc: ["partnerships@clazar.io", psm_if_available]
  })
  gmail.createDraft(draft)
  for task of am_deals:
    hubspot.updateTask(task, "COMPLETED")

HubSpot MCP Slack MCP Gmail MCP Fireflies

Deal Lost Notify

Email

Processes "Co-sell Lost" tasks and creates empathetic Gmail drafts that acknowledge the loss, preserve the relationship, and keep the door open — with a 5-point tone checklist on every draft.

Handles two task formats: Slack link tasks (standard) and inline HTML content tasks (revival motions). Groups by AM, personalizes with engagement history, appends cached Gmail signature, and runs a tone checklist: collegial not vendor-y, acknowledges loss without negativity, soft CTA, "we" language, signature included. Includes closed_lost_reason and notes in composition context.

Before Carefully compose each lost deal email to maintain the relationship. 15+ min per AM (emotionally draining work).

After Tone-checked drafts that sound human, not automated. Review and send. 1 min per AM.

1Load Gmail signature from cached HTML file

2Pull NOT_STARTED "Lost" tasks, fetch associated deals with closed_lost_reason

3Fetch Slack content (link or inline HTML), group by AM

4Research personalization, compose with tone checklist, append signature

5Create Gmail draft, mark tasks COMPLETED

HubSpot MCP Slack MCP Gmail MCP Fireflies

AWS AM Outreach

Email

Intelligently selects from 5 proven email templates based on engagement tier, deal activity, and territory data — each template modeled on actual emails that received replies.

For each AWS AM: builds an intelligence profile (engagement tier, Not Greenfield accounts, active deals, region/segment), then uses a decision tree to pick the right scenario. Active deal? → Co-sell Alignment. Can name specific companies? → Account Followup. International? → Funding/Territory. Prior engagement? → Planning Sync. Cold? → PPA Propensity Report. Every email positions Clazar as a partnerships colleague, not a vendor.

Before Research each AM, figure out the right angle, write a personalized email from scratch. 20+ min per AM.

After AI selects the proven template, personalizes with real data, drafts with correct tone. 2 min per AM.

1Fetch AM contacts from HubSpot, build intelligence profile per AM

2Score engagement tier: Hot (<30d) / Warm (30-90d) / Cold (90d+)

3Decision tree selects scenario: PPA_PROPENSITY | ACCOUNT_FOLLOWUP | FUNDING | COSELL | PLANNING

4Apply template with real data: companies, deals, territory signals

5Run tone philosophy check, create Gmail draft with booking link

// Decision tree for scenario selection
function selectScenario(am) {
  if (am.active_deals.length > 0)
    return "COSELL_ALIGNMENT"
  if (am.not_greenfield_accounts.length >= 2)
    return "ACCOUNT_FOLLOWUP"   // name specific companies
  if (am.region === "international")
    return "FUNDING_TERRITORY"  // $20K AWS funding angle
  if (am.engagement_tier !== "Cold")
    return "PLANNING_SYNC"     // warm contact, request account list
  return "PPA_PROPENSITY"      // highest cold conversion rate
}

// Tone philosophy (enforced on every draft)
// - Position as partnerships colleague, not vendor
// - "We're seeing X on your patch" not "Clazar helps ISVs"
// - Every email offers concrete deliverable
// - Meeting day awareness: Mon→"Tue/Wed?", Thu→"Fri/Mon?"

HubSpot MCP Gmail MCP Fireflies Slack

Email Responder

Email

Scans an incoming email, optionally pulls context from Gmail threads, HubSpot deals, and Slack — then drafts a reply that matches the sender's exact tone and addresses every point.

Parses the email for key points, tone, and action items. Only pulls additional context when it would add value — self-contained emails skip research entirely. Supports iterative refinement (shorter, longer, more formal, custom edits) without re-explaining the full context. Creates a Gmail draft or copies text, user's choice.

Before Read email, look up context in 3 systems, draft reply matching the right tone. 10+ min for complex threads.

After Paste the email → contextual draft with tone matching. Iterate until perfect. 2 min.

1Parse incoming email: sender, subject, key points, tone, action items

2Optionally pull context: Gmail thread history, HubSpot contact/deal, Slack

3Draft reply matching tone, addressing all points

4Present for review → iterate or create Gmail draft

Gmail MCP HubSpot MCP Slack MCP

Cleanup Pending Emails

Auto

Resolves Clay email enrichment tasks that didn't complete during partner cleanup — polls for completions, shows a diff table for confirmation, then updates HubSpot contacts.

Reads the pending tasks file, checks each Clay task's status, extracts completed email addresses, shows before/after changes for user confirmation, and updates HubSpot. Automatically expires tasks older than 24 hours and rewrites the pending file with only in-progress tasks.

Before Manually check each Clay task, copy emails back to HubSpot. Easy to forget pending tasks entirely.

After One command polls all pending tasks, shows diff, updates HubSpot. Expired tasks auto-cleaned.

1Read pending-clay-tasks.json, poll each task via Clay MCP

2Extract completed emails, expire tasks >24h old

3Show diff table for confirmation, update HubSpot contacts

4Rewrite pending file with only in-progress tasks

Clay MCP HubSpot MCP File I/O

CRM & Data Operations4 skills

Keep the CRM clean, the pipeline visible, and the data fresh

Partner Account Mapper

CRM

Maps companies to partner contacts with association labels via a 3-phase process: clear stale associations, clean conflicts, then create new ones — preventing the duplicate and orphan associations that plague manual CRM work.

Supports two CSV formats: full (with company IDs and partner record IDs) or minimal (just company names — auto-enriches domains via HubSpot search + DNS probing across .com/.ai/.io/.co, resolves partner by name). Always runs a dry-run preview first. Groups by partner+label for independent processing. Auto-creates missing companies when domain is resolved.

Before Manual HubSpot associations: find company, find partner, set label, check for conflicts. Error-prone at scale.

After CSV in → dry-run preview → confirm → 3-phase execution with conflict resolution. Minutes for hundreds.

1Parse CSV (detect full vs. minimal format), resolve partner by name if needed

2For minimal CSV: search HubSpot + probe DNS to resolve company domains

3Dry-run preview: show planned changes grouped by partner+label

4Phase 1: Clear stale associations → Phase 2: Clean conflicts → Phase 3: Create new

// Minimal CSV enrichment pipeline
for company_name of csv_rows:
  // Try HubSpot search first
  match = hubspot.search("companies", company_name)
  if (!match) {
    // DNS probe: try common TLDs
    for tld of [".com", ".ai", ".io", ".co"]:
      domain = slugify(company_name) + tld
      if (dnsResolves(domain))
        match = hubspot.create("company", { domain })
  }

// 3-phase execution (prevents duplicates)
for [partner, label, companies] of groups:
  phase1_clear(partner, label)   // remove old
  phase2_clean(partner, label)   // resolve conflicts
  phase3_create(partner, label, companies) // add new

Node.js Script HubSpot REST API DNS Probing Keychain

Partner Contact Cleanup

CRM

Validates every partner contact against their LinkedIn profile, detects job changes and title updates, re-maps company associations, and clears lifecycle stages — keeping the partner database trustworthy.

Runs as a Python script for efficiency. For each contact (processed one-by-one, never batched): scrapes LinkedIn via Apify, compares title and company to HubSpot, applies company equivalences (AWS = Amazon = Amazon Web Services), updates fields, re-maps associations for company movers, and logs company-changed contacts for downstream Clay email enrichment. Fully resumable via progress file after interruptions.

Before Partner data silently goes stale. People change jobs, emails bounce, associations point to wrong companies.

After Every contact validated against LinkedIn. Stale data detected and fixed. ~$6 for 1,500 contacts.

1Pull partner contacts (lifecycle = partner, has high-fit accounts)

2For each: scrape LinkedIn via Apify, compare title + company

3Same company, different title → update. Different company → full re-mapping workflow

4Save progress after each contact (resume with --resume)

Python 3 Apify (LinkedIn) HubSpot REST API Custom Object API

Deal Stage Slack Notifier

Auto

Daily pipeline hygiene: classifies every deal by meeting status, flags stale deals, surfaces upcoming meetings, and sends a consolidated Slack thread grouped by deal owner with specific actions.

Fetches deals in a pipeline stage, classifies each into PAST_MEETING (needs SQL reason update), UPCOMING_MEETING, STALE_NO_MEETING, MISSING_SQL, or CUSTOMER_DELAY. Deduplicates so each deal appears once in its most relevant section. Groups by owner with Slack @mentions for accountability. Skips deals where latest_meeting_date is today (only flags strictly past meetings).

Before Weekly pipeline review: open each deal, check meeting history, ping owners individually. Hours of work.

After Daily automated Slack thread: every deal classified, owners tagged, actions clear. Zero manual review.

1Fetch deals in pipeline stage with latest_meeting_date and partnership__sql_reason

2Classify: PAST_MEETING | UPCOMING | STALE | MISSING_SQL | CUSTOMER_DELAY

3Deduplicate (one deal = one flag, most relevant wins)

4Group by owner, send parent Slack message + threaded sections

HubSpot MCP Slack MCP

Daily AWS Call Notes

Auto

Automatically processes today's partner calls: extracts every company mentioned, matches to HubSpot, verifies AM associations via direct API, and writes structured notes — so nothing from a call is ever lost.

Fetches today's transcripts from Fireflies, extracts company names from attendees, title, and content (with equivalences: AWS/Amazon/Google Cloud/Azure treated as same). Cross-references to HubSpot companies, verifies Account Manager partner contact associations via custom object API (associationTypeId 580), and creates formatted notes with AM name, attendees, duration, summary bullets, and action items. Always reports unmatched companies so nothing falls through the cracks.

Before Listen to call recordings, identify companies, find them in HubSpot, write up notes manually. 30+ min per call.

After Runs automatically. Notes appear on HubSpot company records. Unmatched companies flagged. 0 manual work.

1Fetch today's transcripts from Fireflies

2Extract company names (attendees, title, content), apply equivalences

3Match to HubSpot, verify AM partner contact associations via direct API

4Write structured notes (check for existing today's note first), report summary

Fireflies MCP HubSpot MCP Custom Object API Keychain

Infrastructure1 skill

The meta-skill that makes all other skills run on schedule

Skill Scheduler

Infra

Generates macOS LaunchAgent plists + launcher scripts for any Claude Code skill, turning one-off commands into scheduled automations that run daily, hourly, or at specific times — with lock files, weekday checks, and binary caching.

Given a skill name and schedule expression, generates two files: a bash launcher script (with lock file deduplication, weekday enforcement, Claude Code binary resolution + caching, and Terminal.app launch via AppleScript) and a LaunchAgent plist for macOS daemon scheduling. Supports "daily at 9am", "every 4 hours", or specific time formats. All skills run via claude -p "/skill" with --allowedTools '*' for headless execution.

Before Remember to run each skill manually. Miss a day? Pipeline review doesn't happen. Call notes aren't captured.

After /scheduler deal-notify "daily at 9am" → runs forever, skips weekends, never overlaps.

1Parse skill name + schedule format (daily / every Xh / at HH:MM)

2Generate launcher script with lock file, weekday check, binary caching

3Generate LaunchAgent plist (~/Library/LaunchAgents/)

4Output launchctl load command for activation

// Generated launcher script (simplified)
#!/bin/bash
LOCK="/tmp/claude-skill-{name}.lock"

# Skip weekends
[[ $(date +%u) -gt 5 ]] && exit 0

# Lock file prevents overlapping runs
[[ -f "$LOCK" ]] && exit 0
touch "$LOCK"

# Resolve + cache Claude Code binary
CLAUDE=$(which claude || find /usr/local -name claude)
echo "$CLAUDE" > /tmp/.claude-bin-cache

# Launch in Terminal.app via AppleScript
osascript -e "tell app \"Terminal\" to do script
  \"$CLAUDE -p '/{name}' --allowedTools '*'\""

rm "$LOCK"

Bash LaunchAgent AppleScript Terminal.app

Compound Workflows

Skills that chain together

Individual skills are useful. But the real power comes from how they compose — the output of one skill becomes the input context for another.

Partner Onboarding Pipeline

New partner assigned → account mapper creates associations → partner intelligence builds their Notion profile → team lookup maps their org → AM outreach drafts the first email.

Account Mapper → Partner Intel → Team Lookup → AM Outreach

Daily Operations Loop

Scheduled daily: call notes capture today's meetings, deal notifier flags pipeline issues, co-sell drafter processes pending tasks, cleanup resolves pending enrichments.

Call Notes → Deal Notify → Co-Sell Drafter → Cleanup Emails

Data Hygiene Cycle

Partner cleanup validates contacts against LinkedIn → company movers get re-mapped → pending email enrichments are resolved by cleanup skill → intelligence pages refresh.

Contact Cleanup → Account Mapper → Cleanup Emails → Partner Intel

Prospect Response Workflow

Prospect replies → bridge intel builds AM context package with social proof → email responder drafts the reply → AM outreach updates the partner engagement.

Bridge Intel → Email Responder → AM Outreach

Engineering Principles

Patterns that make it reliable

Every skill follows the same core patterns. These aren't theoretical — each was learned from a production failure or a workflow bottleneck.

⚡

Parallel Execution

Independent data fetches launch in a single Claude message block. Partner Intelligence fetches from 6 sources simultaneously — HubSpot, Fireflies, Slack, Notion, Web, Custom Objects — in one round trip. Cuts latency from sequential minutes to a single parallel batch.

USED IN: Partner Intel, Co-Sell Drafter, Bridge Intel

🔄

Resumable Processing

Long-running skills save progress after each item. Partner Contact Cleanup writes cursor position to a JSON file after every contact. If context exhausts or connection drops, --resume picks up exactly where it left off. No duplicate work, no missed contacts.

USED IN: Contact Cleanup, Cleanup Pending Emails

🔗

Dual Token Architecture

HubSpot's MCP connector doesn't support custom objects. So the system uses two tokens: standard HubSpot token for contacts/deals/tasks via MCP, and a custom object token for Partner Contact (2-40826561) associations via direct REST API. Both stored in macOS Keychain.

USED IN: Partner Intel, Team Lookup, Call Notes, Contact Cleanup

🛡️

Never Fabricate, Never Block

If a data source is empty, the skill marks it as "—" and continues. Missing Fireflies transcripts don't block partner intelligence from writing the other 5 sections. Missing AM email gets flagged for manual review. Partial intel always beats complete failure.

USED IN: All 13 skills

📧

Human-in-the-Loop Drafts

Email skills create Gmail drafts, never send directly. Co-sell emails, lost deal notifications, AM outreach — all land in your drafts folder for review. The AI handles research, personalization, and composition. The human handles the send button.

USED IN: Co-Sell Drafter, Deal Lost, AM Outreach, Email Responder

🔍

Company Equivalences

"Amazon", "AWS", and "Amazon Web Services" are the same company. Same for Google/Google Cloud, Microsoft/Azure. Every skill that matches company names applies equivalence rules to prevent false negatives. Critical for call notes, contact cleanup, and partner intelligence.

USED IN: Call Notes, Contact Cleanup, Partner Intel

📊

Deduplication at Every Layer

Deal Notifier: each deal appears once in its most relevant section. Account Mapper: 3-phase process prevents duplicate associations. Co-Sell Drafter: multi-deal emails combine into one per AM. Contact Cleanup: progress file prevents re-processing. Every output is deduplicated.

USED IN: Deal Notify, Account Mapper, Co-Sell Drafter, Cleanup

⏰

Meeting Day Awareness

Email skills adjust CTA language based on the current day of the week. Monday → "perhaps Tuesday or Wednesday?" Thursday → "perhaps Friday or Monday?" A small detail that makes automated emails feel human-written.

USED IN: Co-Sell Drafter, AM Outreach, Deal Lost

Impact

What this adds up to

Partner Profile Build

~2 min

was 45+ minutes per profile

Co-Sell Email Turnaround

Task → Draft

end-to-end, no manual steps

Daily Call Notes

Automated

extracted, matched, verified, written

Pipeline Review

Daily

classified + Slack thread by owner

Contact Validation

~$6 / 1,500

LinkedIn-verified, resumable

Manual Data Entry

Zero

CSV in, associations out

I replaced 40 hours of manual work with 13 AI automations