Nooks Autopilot

Nooks Autopilot — Autonomous BDR/SDR Team

Per-rep agents. This runs as a dedicated agent for each dialing rep (Tim + the 3 SDRs). Each rep's run works only that rep's owned leads and produces their next-best top 5 to call. One owner per lead, no double-dialing: every worked lead is owned by exactly one rep; a lead is never enrolled twice or placed in two reps' queues. Unowned source-list leads are assigned to exactly one rep (whole account → one rep, to keep multi-threading clean) before anything else.

SHADOW MODE is the default and is ON unless explicitly turned off. In shadow mode the skill does everything except write to Nooks/HubSpot — it produces an enrollment plan + report so Tim can verify behavior on real leads before a single sequence starts.

SHADOW_MODE: true (master switch — no Nooks/HubSpot writes while true) DAILY_ENROLL_CAP: 25 (throttle once live; protects deliverability) ENROLLMENT GATE (auto): persona ∈ focus verticals AND power_level == atl AND not junk AND survived Golden Rules / Suppression / dedup. The qualify_lead CatBoost category is NOT a gate — it's a ranking input (a focus-vertical ATL with a "why now" signal still enrolls even if category=nurture; the classifier under-rates Live Events). GRAY power → hold for Tim's ok; BTL → skip auto-enroll. TOP_N_PER_REP: 5 (each rep's "next best to call" queue size)

Dialing reps (one owner per lead; Nooks crmId == HubSpot ownerId; mailbox = auto-email send-as):

(Victor Doubrovine — manager, ownerId 153471050 / Nooks D9DElZ5… — is NOT a dialing rep; never assign to him.)

Ownership key = sdr_owner (HubSpot enum of owner-ids; Victor enabled it and it translates into Nooks as the dial owner). This — NOT hubspot_owner_id (often the AE) — is the per-rep ownership field. A lead "belongs to" the rep in its sdr_owner.

Assignment strategy: assign by account (all of an account's contacts → one rep) to prevent two reps working the same org and to keep multi-threading owned by one rep. New sdr_owner-unset accounts → round-robin/load-balanced across the 4 reps; if any contact at the account already has an sdr_owner, assign the rest to that same rep. Never overwrite an existing sdr_owner. Set sdr_owner (live) via hubspot_update_contact — it propagates to Nooks as the owner.

Focus verticals (v1): Higher Ed, Community College, Live Events, Corporate. Other personas stay matchable but are deprioritized.

Source lists (HubSpot) → play:

State lists (HubSpot, DYNAMIC) — driven by the epiphan_ai_status contact property (enum):

To reflect state: write epiphan_ai_status on the contact (do NOT edit the dynamic list). Related mirror props: nooks_active_in_sequence (bool — double-dial guard), nooks_sequenced_by (owning rep), nooks_current_sequence_name, nooks_current_sequence_step, epiphan_ai_last_email_at.

Warm signals (Stage 7): epiphan_ai_status == replied (list 4576) OR a Nooks call disposition of Meeting Booked / Connected / Referral. (Negative: No Interest, No Longer at Company, Wrong number.)

Routing → Nooks sequence id:

Default when ambiguous → General Prospecting. Signals (job-change, lost-deal) override the list→play default.

Stage 1 — Source

For each configured source list, call hubspot_list_members(listId, gapsOnly:false). Capture name, email, jobtitle, company, phone, contactId, plus the per-member gap flags. Tag each lead with its play (from the Source-lists table). Cap the run at DAILY_ENROLL_CAP candidates (highest-intent lists first: Hit Lists → Greenfield → Send-to-AI).

Stage 2 — Filter (reuse existing logic)

Apply, in order — these mirror sequence-load Stage 3 + prospect-research-to-cadence Stage 1b:

• Golden Rules: skip customers, channel partners, device-owners, contacts owned by active AEs (<90 days), and NEVER-ATL titles (Warehouse/Network Manager, Systems Admin, AV Technician, Classroom Support, Lab Coordinator, Multimedia Services Manager, Video Production Specialist, etc.).
• Suppression Gate: exclude where bdr_suppression_until is set and > today.
• Dedup: drop anyone already in [Epiphan AI] Enrolled (4575) or [Epiphan AI] Blocked (4577).

Stage 2b — Ownership + no-double-dial guard (CRITICAL)

A lead must be worked by exactly one rep and never dialed by two. Enforce, in order:

1. Already in motion (double-dial guard): exclude any candidate with nooks_active_in_sequence == true (fast HubSpot filter; nooks_sequenced_by shows which rep). Cross-check authoritatively against Nooks listSequenceStates(filter_state:[active,paused], include:[prospect]) on prospect.crmId == contactId. Also exclude epiphan_ai_status ∈ {enrolled, sent, replied} (already in the AI pipeline).
2. Existing SDR owner: read sdr_owner.
   • Set to a dialing rep → belongs to that rep's queue only (never another rep's).
   • Set to an AE/non-dialer → leave per Golden Rules (stale-AE exception handled upstream).
   • Unset → eligible for assignment.
3. Assign unset by account: group eligible candidates by account/domain; assign the whole account to one rep (round-robin/load-balanced; if any contact at the account already has an sdr_owner, use that rep). SHADOW: propose the assignment in the report. LIVE: set sdr_owner via hubspot_update_contact (propagates to Nooks). Never overwrite an existing sdr_owner.
4. Result: a clean per-rep candidate set with zero cross-rep overlap.

Stage 2c — DA Audit (always-on AE-pipeline + recent-purchase guard) (CRITICAL)

A devil's-advocate pass that must run on every candidate account before enrollment — challenge each one with "is there a reason we should NOT touch this?". Two hard exclusions:

1. Active AE deal (Lex / Phil): if the account has an open deal owned by or collaborated on by Lex Evans (82625923) or Phil Sandler (190030668) → EXCLUDE (do not autopilot-dial an account with a live AE opportunity). Surface it as AE-ACTIVE → coordinate with <AE>, never enroll. Check via hubspot_search_deals (owner/collaborator filter, open stages) associated to the account, or Epiphan AI ask_agent / sales_brief for the account's open pipeline.
2. Recent purchase (last 12 months): if the account purchased within the last 12 months (crm_get_customer_orders / query_dataset revenue / closed-won deal in window) → EXCLUDE as a recent customer; route to CS/expansion, not cold dial. (Catches recent buyers that a stale lifecyclestage would miss.) Run this as a real audit step (optionally via a devils-advocate subagent over the candidate list); log every exclusion with its reason. Also re-applies the standard customer/channel checks as backstop.

Stage 3 — Qualify + score

• qualify_lead (Epiphan AI) → category (ideal_mql/warm/nurture/junk), power_level (atl/btl/unknown), region, junk flag. Drop junk; nurture → leave for email-only, don't auto-enroll the dial cadence.
• Persona match to a v1 vertical persona using the exact title filters in reference/nooks-personas-v2.md (Live Event Producer, Higher Ed Lecture Capture Lead, Community College Media Coordinator, Corporate Video Specialist). Respect each persona's "Title does NOT include" exclusions.
• Win-pattern weighting (from Tim's + AE conversion data, see reference/): boost AV/Media Services Manager–Director and Head/Director of Instructional Technology (incl. Law/Med variants); down-weight central-IT-only with no AV project and junior specialists.
• "Why now" signal per reference/nooks-signals-v1.md: legacy-capture-HW EOL (Matrox/Extron SMP), competing capture/PC-Windows tech, job-change, recent expansion. For v1, inherit the signal from the source-list context (e.g. a Greenfield/EOL-bucketed list) and any obvious title/firmographic cue; full detection is v3.
• Auto-enroll gate (see config): persona-in-focus + power_level == atl + not junk → eligible. Category nurture does not block (it under-rates Live Events). GRAY → hold for Tim; BTL → skip.
• Priority score (drives per-rep ranking + the top-5; category is one input, not a gate): score = 0.30·signal_strength + 0.25·ICP_vertical_score + 0.20·win_pattern_weight + 0.15·category_rank(ideal_mql=1.0 / warm=0.6 / nurture=0.3) + 0.10·phone_present. Signal strength: legacy-HW-EOL / competing-tech / job-change = high; expansion/hiring = medium; none = low. Win-pattern: AV/Media Manager–Director + Head/Dir Instructional Technology = high; central-IT-only / junior = low.

Stage 4 — Route

Map play → Nooks sequence id (Routing table). A detected job-change or lost-deal signal overrides the list default. Record sequence_id, sequence_name, and a one-line routing_reason.

Stage 5 — Enroll (SHADOW-GATED)

If SHADOW_MODE == true (default): DO NOT call any Nooks/HubSpot write tool. Build the enrollment plan and emit the report (Stage 8). Each row: lead, persona, signal, ATL/BTL, phone status, routed sequence + reason, and what would happen ("would enroll", "would set [Epiphan AI] Enrolled").

If SHADOW_MODE == false: per ATL lead, up to DAILY_ENROLL_CAP:

1. Resolve the Nooks prospect: listSequenceStates/prospect search by email; the Nooks prospect.crmId equals the HubSpot contactId. If absent, create the prospect (syncProspects/create).
2. createSequenceState({data:{prospect:{id}, sequence:{id}, owner:{id: <rep Nooks userId>}, mailbox:{id: <rep mailbox>}, state:"active"}}). The owner (rep's Nooks userId, mapped from sdr_owner via the roster table) makes the cadence + its call/manual tasks land in that rep's queue; the mailbox sends auto-emails from that rep. (Target sequences are team_editable, so any owner is allowed.)
3. Reflect state on the HubSpot contact: set epiphan_ai_status = "enrolled" (drives [Epiphan AI] Enrolled 4575). The nooks_active_in_sequence / nooks_sequenced_by / nooks_current_sequence_name mirrors are maintained by the Nooks↔HubSpot sync — confirm at first live run; write them if the sync lags.
4. Respect the per-rep daily cap; log any failures and continue (graceful degradation).

Stage 6 — Multi-thread, single-owner (CRITICAL guard)

Multi-threading is only safe if one SDR owns the entire account. Never let 2–3 SDRs dial the same company — that's worse than not multi-threading.

• Account = atomic ownership unit. Every contact at an account (grouped by registrable domain) carries the same sdr_owner. When the skill adds a 2nd/3rd contact (multi-thread), that contact inherits the account's existing sdr_owner — never round-robin a teammate onto an owned account.
• Conflict detection: if an account already has contacts with different sdr_owner values, flag it as a SPLIT ACCOUNT and consolidate to one rep (default: the rep who owns the most/most-senior contact, or who has an active Nooks enrollment there) before enrolling anyone new.
• Missing economic buyer: if an account has no Manager/Director-level (ATL) contact in the plan, call hubspot_find_contacts_by_role(domain, titles:[AV/Media/IT Director, Head of Instructional Technology]) to surface it — and attach it to the same sdr_owner as the rest of the account. Rationale (evidence): every 2026 single-threaded deal below Manager level lost.

Stage 7 — Monitor + warm handoff (runs every invocation)

Detect warm prospects via two signals (no email-body parsing needed):

• Reply: epiphan_ai_status == replied (membership in [Epiphan AI] Replied 4576) — the Epiphan AI pipeline already detects email replies.
• Call outcome: a Nooks call disposition of Meeting Booked / Connected / Referral on the prospect (poll listCalls/dispositions, or listSequenceStates(filter_updatedAt_gte: last_run) for states that changed). A finished state with none of these = cold-complete (status → completed), not warm.

For each warm prospect:

• finishSequenceState(id) if still active/paused (halt remaining touches).
• Set epiphan_ai_status = "replied" on the contact.
• Notify the owning rep (sdr_owner): hubspot_create_contact_task(contactId, ownerId: <sdr_owner>, taskType:CALL, subject:"WARM — replied/booked in Nooks", dueDate: today) and surface in that rep's next morning-brief.

Stage 7b — Per-rep "next best 5 to call"

Prioritize each rep's owned + newly-assigned candidates and surface their TOP_N_PER_REP (5). Rank by: persona-fit + ICP vertical score, "why now" signal strength (EOL/competing-tech/job-change highest), ATL power level, win-pattern weighting (AV/Media Manager–Director + Instructional Technology up; central-IT-only/junior down), and phone-present. Each entry: name, title, company, phone, persona, signal, routed sequence, one-line "why call now". This is the rep's dial list for the block — and it contains only leads no other rep is working (Stage 2b guarantees it).

Stage 8 — Report (per rep) + outcome sidecar

Fill reference/dial-list-report-template.html (self-contained; repeat the .rep block per dialing rep, the .card block ×5 for their Next-Best-5, and a .candidates row per lead) and write it to the output dir as nooks_autopilot_<YYYY-MM-DD>.html. Mode badge = SHADOW or LIVE. Each rep section shows their Next-Best 5 to Call (ranked by the priority score, with a one-line "why now") + a collapsible full candidate table (disposition: would-enroll / enrolled / skipped+reason, persona, signal, sequence, ATL/BTL, phone, multi-thread gaps). Header KPIs: candidates, qualified ATL, planned/enrolled, double-dials prevented, warm handoffs. Then write ~/.claude/skill-analytics/last-outcome-nooks-autopilot.json:

{"ts":"[UTC ISO8601]","skill":"nooks-autopilot","version":"1.0.0","variant":"[shadow|live]",
 "status":"[success|partial|error]","runtime_ms":[ms],
 "metrics":{"candidates":[n],"qualified_atl":[n],"shadow_planned":[n],"enrolled":[n],
            "skipped_golden_rules":[n],"skipped_suppressed":[n],"skipped_dedup":[n],
            "double_dials_prevented":[n],"accounts_assigned":[n],
            "multithread_gaps":[n],"warm_handoffs":[n],
            "per_rep":{"<ownerId>":{"top5":[n],"enrolled":[n],"assigned":[n]}}},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Sibling skills referenced (reuse, don't rebuild)

• sequence-load — enrollment scaffold, Golden Rules, dedup, phone validation, suppression gate.
• prospect-research-to-cadence — qualify/score, ATL/BTL classification, ICP scoring, Clay waterfall.
• prospect-enrich / phone-verification-waterfall — phone enrichment.
• morning-brief — surface warm handoffs to Tim.

Reference data (this skill's reference/)

nooks-personas-v2.md (title filters per persona), nooks-signals-v1.md (why-now triggers), nooks-battlecards-v3.md (killer-Q arcs for v3 content personalization), Higher_Ed_Survey_Prospect_Signals_Report.md (P0/P1 named accounts). Drive canon: SDR System Map, Prompt Playbook, Sales Stack Guide (the operating model).

Build-time discovery — RESOLVED (2026-06-19)

1. ✅ Controlling property = epiphan_ai_status (enum: enrolled/sent/replied/completed/skipped_*). Ownership key = sdr_owner (translates to Nooks owner). Double-dial guard = nooks_active_in_sequence.
2. ✅ Enroll attribution: createSequenceState takes owner.id (rep Nooks userId) + mailbox.id.
3. ✅ Warm detection: epiphan_ai_status==replied OR Nooks disposition Meeting Booked/Connected/Referral.

Remaining confirmations (first live run only)

• Each rep's Nooks mailbox id (listMailboxes filter by user) for auto-email send-as.
• Whether the Nooks↔HubSpot sync auto-maintains nooks_active_in_sequence/epiphan_ai_status on a Nooks-side enrollment, or this skill must write them.
• Coordinate with the existing Epiphan AI prospecting agent (it also sets epiphan_ai_status and may enroll via Apollo) so a contact isn't worked by both pipelines — treat epiphan_ai_status ∈ {enrolled,sent,replied} and hs_currently_enrolled_in_prospecting_agent=true as "hands off".

Guardrails

• SHADOW first. Never make a Nooks/HubSpot write while SHADOW_MODE == true.
• Auto-enroll ATL only; GRAY needs Tim's ok; never auto-enroll BTL.
• Honor the daily cap and the Suppression Gate every run.
• Validate any customer-facing claim via Epiphan AI product search / the Verified spec bank before content personalization (v3) — never quote specs from memory.
• The agent only mutates [AI] -prefixed lists it owns; state is reflected via contact properties.

Scheduling

• Cron: schedule frontmatter 0 5 8 * * 1-5 (sec-min-hour order, matching sibling skills) = weekdays 08:05 America/New_York, right after morning-brief.
• One run, all reps: a single daily run produces the report segmented per rep (each rep reads their own section + Next-Best-5). Simpler than 4 crons and gives every SDR their queue. (Optional later: split into per-rep scheduled runs if reps want independent timing.)
• Monitor pass (Stage 7) runs on every invocation; optionally add a lighter midday cron (0 0 13 * * 1-5) so same-day replies/bookings get halted + handed off quickly.

Skill metadata

Version: 1.0 · Author: Tim Kipper · Status: v1 shadow · Tier: P1 (Core BDR Automation) Chain: prospect-refresh → sequence-load → morning-brief → nooks-autopilot Supersedes: the FastAPI backend autonomous pipeline for the live BDR/SDR motion.