gtm-brain-skill

Graph lives at: neo4j+s://23a749c7.databases.neo4j.io (AuraDB Free — GTM Knowledge Graph) Credentials: skills/.env (never hardcode elsewhere) Schema reference: reference/graph-schema.md Query patterns: reference/cypher-patterns.md </objective>

<quick_start> Setup (run once per machine):

pip3 install neo4j
set -a && source ~/Desktop/tk_projects/skills/.env && set +a
python3 ~/.claude/skills/gtm-brain-skill/scripts/brain.py init

Test connection:

set -a && source ~/Desktop/tk_projects/skills/.env && set +a
python3 ~/.claude/skills/gtm-brain-skill/scripts/brain.py ping

Common intents → Stage to jump to:

<success_criteria>

• Connection ping returns {"ok": 1, "msg": "GTM Brain connected"}
• READ queries return ranked results with vertical/title/outcome breakdown
• WRITE operations confirm node upsert and relationship creation
• No credentials in any file other than skills/.env
• Outcome sidecar written after each session </success_criteria>

<core_content>

ENV Helper (prepend to every bash block)

set -a && source ~/Desktop/tk_projects/skills/.env && set +a
SCRIPT=~/.claude/skills/gtm-brain-skill/scripts/brain.py
# fallback to repo path if deployed path missing:
[ -f "$SCRIPT" ] || SCRIPT=~/Desktop/tk_projects/skills/active/gtm-brain-skill/scripts/brain.py

Stage 1 — Connect

Always ping first to confirm Aura is awake (free instances sleep after inactivity):

set -a && source ~/Desktop/tk_projects/skills/.env && set +a
SCRIPT=~/.claude/skills/gtm-brain-skill/scripts/brain.py
[ -f "$SCRIPT" ] || SCRIPT=~/Desktop/tk_projects/skills/active/gtm-brain-skill/scripts/brain.py
python3 "$SCRIPT" ping

Expected: {"ok": 1, "msg": "GTM Brain connected"}

If it hangs 60+ seconds → Aura instance was sleeping. Wait 60 seconds and retry.

Stage 2 — Intent Routing

Classify what the user wants:

Stage 3 — READ Queries

Construct the correct Cypher from reference/cypher-patterns.md, then run:

python3 "$SCRIPT" run "<CYPHER_QUERY>"

Pattern: What worked with vertical + title?

python3 "$SCRIPT" run "
MATCH (c:Contact {vertical: 'Higher Ed', atl_btl: 'ATL'})-[:HAD]->(o:Outcome {result: 'positive'})
RETURN c.title, o.channel, o.notes, count(*) AS wins
ORDER BY wins DESC LIMIT 10"

Format output as a ranked table: Title | Channel | Win count | Best notes sample.

Pattern: Account map — all contacts + outcomes

python3 "$SCRIPT" run "
MATCH (c:Contact)-[:WORKS_AT]->(a:Account)
WHERE toLower(a.name) CONTAINS toLower('[COMPANY]')
OPTIONAL MATCH (c)-[:HAD]->(o:Outcome)
RETURN c.name, c.title, c.atl_btl,
       collect({ch: o.channel, r: o.result, n: o.notes}) AS history
ORDER BY c.atl_btl"

Format as: Name | Title | Tier | Last outcome.

Pattern: Sequence performance by vertical

python3 "$SCRIPT" run "
MATCH (c:Contact {vertical: '[VERTICAL]'})-[:HAD]->(o:Outcome {result: 'positive'})-[:VIA]->(s:Sequence)
RETURN s.name, count(o) AS wins ORDER BY wins DESC LIMIT 5"

Pattern: ATL coverage on open deals

python3 "$SCRIPT" run "
MATCH (d:Deal) WHERE d.stage NOT IN ['Closed Won','Closed Lost']
MATCH (c:Contact)-[:INVOLVED_IN]->(d)
WITH d, collect(c.atl_btl) AS tiers
RETURN d.name, d.stage, d.amount,
       size([t IN tiers WHERE t='ATL']) AS atl_count, size(tiers) AS total
ORDER BY atl_count ASC"

Flag ⚠️ for deals with 0 ATL contacts.

Stage 4 — WRITE Operations

Log a call or email outcome

When a call/sequence/email completes, write the outcome:

1. Confirm contact exists in graph (merge if not):

python3 "$SCRIPT" merge-contact \
  '{"hubspot_id":"<HS_ID>","name":"<NAME>","title":"<TITLE>","vertical":"<VERTICAL>","atl_btl":"<TIER>","email":"<EMAIL>"}'

2. Log the outcome:

python3 "$SCRIPT" log-outcome \
  '{"id":"outcome-<YYYYMMDD>-<HS_ID>","contact_hubspot_id":"<HS_ID>","channel":"call","result":"<positive|neutral|negative>","notes":"<what happened>","timestamp":"<ISO8601>","sequence_nooks_id":"<ID_IF_IN_SEQ>"}'

Add / update account

python3 "$SCRIPT" merge-account \
  '{"hubspot_id":"<HS_ID>","name":"<NAME>","vertical":"<VERTICAL>","icp_score":<SCORE>}'

Wire contact to deal

python3 "$SCRIPT" run \
  "MATCH (c:Contact {hubspot_id:'<C_ID>'}),(d:Deal {hubspot_id:'<D_ID>'}) \
   MERGE (c)-[:INVOLVED_IN {role:'champion'}]->(d)"

Wire contact to account (if not already set)

python3 "$SCRIPT" run \
  "MATCH (c:Contact {hubspot_id:'<C_ID>'}),(a:Account {hubspot_id:'<A_ID>'}) \
   MERGE (c)-[:WORKS_AT]->(a)"

Stage 5 — Schema Initialization (first time only)

set -a && source ~/Desktop/tk_projects/skills/.env && set +a
python3 "$SCRIPT" init

Creates: uniqueness constraints on all primary keys, indexes on vertical/atl_btl/result. Safe to re-run (all statements use IF NOT EXISTS).

Integration Points

Other skills that should write outcomes to GTM Brain:

• epiphan-call-playbook — after disposition is logged, call Stage 4 to log outcome
• sdr-call-coaching — after scoring a call, log result to graph
• nooks-autopilot — after warm handoff, log positive outcome

Other skills that should read from GTM Brain:

• morning-brief-skill — prepend "who responded before at this vertical" intel
• sdr-dial-lists — surface prior positive contacts at target accounts
• meddic-call-prep-auto-skill — pull contact history before a call

Manual workflow until wired: Say "log outcome to GTM Brain for [contact]" after any call or email to capture the result.

Stage 6 — HubSpot Sync

Pull a contact and their company from HubSpot and write both to the graph.

Trigger: "sync [name] from HubSpot" / "add [company] to graph" / "pull [contact] from HS"

Step 1 — Fetch contact from HubSpot MCP

Use hubspot_search_contacts (Epiphan AI MCP) with the name or email:

mcp__claude_ai_Epiphan_Ai__hubspot_search_contacts(query="Jane Smith")

Extract from result: id (hubspot_id), firstname, lastname, jobtitle, email, phone, associatedcompanyid.

Step 2 — Classify ATL/BTL from title

Apply the ATL/BTL Classification from CLAUDE.md:

• Match title against Universal ATL Keywords → ATL
• Match against BTL / NEVER ATL lists → BTL or NEVER
• Gray zone → GRAY

Step 3 — Determine vertical

Use mcp__claude_ai_Epiphan_Ai__qualify_lead or infer from company name/domain: Higher Ed → university/college; Courts → court/judicial; Government → city/county/state agency; Healthcare → hospital/health system; Corporate AV → enterprise/corporate.

Step 4 — Write contact to graph

python3 "$SCRIPT" merge-contact \
  '{"hubspot_id":"<ID>","name":"<FIRST> <LAST>","title":"<JOBTITLE>","email":"<EMAIL>","phone":"<PHONE>","vertical":"<VERTICAL>","atl_btl":"<TIER>"}'

Step 5 — Fetch and write company

mcp__claude_ai_Epiphan_Ai__hubspot_get_company(companyId="<associatedcompanyid>")

Extract: id, name, domain, vertical (same logic). Write:

python3 "$SCRIPT" merge-account \
  '{"hubspot_id":"<CO_ID>","name":"<CO_NAME>","vertical":"<VERTICAL>","domain":"<DOMAIN>","icp_score":80}'

Wire the relationship:

python3 "$SCRIPT" run \
  "MATCH (c:Contact {hubspot_id:'<C_ID>'}),(a:Account {hubspot_id:'<A_ID>'}) MERGE (c)-[:WORKS_AT]->(a)"

Batch sync tip: "sync everyone at [company] from HubSpot" → hubspot_search_contacts with company filter → loop Steps 2-5 for each contact.

Stage 7 — Nooks Call Sync

Import recent Nooks call dispositions as Outcome nodes. Run after a dial session.

Trigger: "import today's Nooks calls" / "sync calls to graph" / "log calls from Nooks"

Step 1 — List recent calls

mcp__claude_ai_Nooks__listCalls(filter_owner_id="87486452", filter_time_gte="<TODAY_ISO>")

Use Tim's Nooks user ID 87486452. Returns call list with id, prospectId, sequenceId.

Step 2 — Get disposition per call

For each call:

mcp__claude_ai_Nooks__getCallDisposition(callId="<id>")

Step 3 — Map disposition → result

Step 4 — Resolve contact hubspot_id

mcp__claude_ai_Nooks__getProspect(prospectId="<id>")

Extract hubspotContactId. If contact not yet in graph → run Stage 6 sync first.

Step 5 — Write each call as an Outcome

python3 "$SCRIPT" log-outcome \
  '{"id":"nooks-<CALL_ID>","contact_hubspot_id":"<HS_ID>","channel":"call","result":"<positive|neutral|negative>","notes":"<disposition label + any call notes>","timestamp":"<call_start_time>","sequence_nooks_id":"<SEQ_ID_IF_SET>"}'

Process all calls in the batch. Report: N calls imported, breakdown by result (X positive / Y neutral / Z negative).

</core_content>

Emit Outcome Sidecar

Write to ~/.claude/skill-analytics/last-outcome-gtm-brain.json:

{
  "ts": "<ISO-8601>",
  "skill": "gtm-brain",
  "version": "1.0.0",
  "variant": "control",
  "status": "<success|partial|error>",
  "runtime_ms": <int>,
  "metrics": {
    "intent": "<read|write|init|ping>",
    "nodes_written": <int>,
    "rows_returned": <int>,
    "vertical": "<if applicable>"
  },
  "error": null,
  "session_id": "<YYYY-MM-DD>"
}