Grok Bot SDR Guide: Cursor Handles Research and Drafts, Bots Handle Deduplication and Staging Triage

This Grok Bot SDR guide illustrates a two-tool collaborative prospecting workflow. Sales Development Representative (SDR) Stefan Markarian uses Cursor cloud agents to synthesize account research and first-touch emails from Hex, Salesforce, and Gong, then relies on Grok Bot to check cross-system statuses, create records, schedule unsent sequences, and maintain prospecting sheets.

When a contact's status is clear, staging proceeds; when paused states, early-stage opportunities, or identity ambiguities arise, contacts are routed to FLAG for sales discretion. Emails are reviewed and sent by humans; blank LinkedIn connection requests without notes are handled within strict limits by a dedicated John Ford Bot.

Dual-Engine Division of Labor: How the Author Pairs Cursor and Grok Bot

When attempting to overhaul sales workflows with LLMs, the most common mistake teams make is trying to do everything at once—writing a sprawling, monolithic prompt expecting a single AI assistant to look up background data, draft copy, and click buttons across systems to send messages. The outcome is almost always the same: either the AI hallucinates midway and breaks the flow, or it blasts malformatted spam directly to high-priority accounts.

Stefan's architecture centers on: cleanly separating heavy cognitive synthesis and text drafting from cross-system browser legwork, delegating each to the best-suited tool.

┌────────────────────────────────────────────────────────┐
│       Overnight Background: Cursor Synthesis Engine    │
│    Hex (Footprint) + Salesforce (Opp) + Gong (Calls)   │
│                       ↓                                │
│       Output: Structured Account Research Packet       │
│       + First-Touch Icebreaker in Personal Voice       │
└──────────────────────────┬─────────────────────────────┘
                           │ Packet Handoff
┌──────────────────────────▼─────────────────────────────┐
│       Morning Resident: Grok Bot Specialized Matrix    │
│                                                        │
│  ┌────────────────┐  Intake    ┌───────────────────┐   │
│  │   Josh Allen   │ ─────────► │ CLEAR: Pre-Staging│   │
│  │(Chief of Staff)│ Deduplicate│ FLAG : Human Queue│   │
│  └───────┬────────┘            └───────────────────┘   │
│          │ Parallel Delegation                         │
│   ┌──────┴──────────────┐                              │
│   ▼                     ▼                              │
│ Contact Man (Sheets)    John Ford (LinkedIn)           │
└──────────────────────────┬─────────────────────────────┘
                           │ Summarize
┌──────────────────────────▼─────────────────────────────┐
│         Morning Shift: Human SDR Command Console       │
│  1. Quickly review 1 uncontroversial CLEAR email & send│
│  2. Focus cognitive effort on 34 complex FLAG cases    │
│  3. Emails sent by human; blank LinkedIn connection    │
│     requests processed by John Ford within limits      │
└────────────────────────────────────────────────────────┘

1. Cursor: The "Offline Brain" for Deep Research and Personalized Drafting

• Stage Ownership: Research, Prospecting, and Drafting.
• Core Strength: Built on a local or cloud code-editor-grade workflow, Cursor excels at long-context synthesis. Guided by specialized sales research skills, it dives directly into Hex to pull product telemetry, digs through Salesforce for closed-lost history, and distills Gong call recordings to assemble a structured "Account Research Packet"—drafting the first-touch icebreaker in the SDR's authentic voice before morning.
• Deliverable: Not an automatically dispatched email, but a high-quality, evidence-backed draft ready for review.

2. Grok Bot: The "Always-On Field Engineer" for Cross-System Plumbing and Intake Triage

• Stage Ownership: System Intake, Multi-Platform Deduplication, Status Synchronization, and Cross-Tool Execution.
• Core Strength: Runs as a persistent resident process, continuing to execute even when your laptop lid is closed. It ingests Cursor's outputs, relieving humans from manually cross-checking records across Nooks, Amplemarket, Salesforce, and LinkedIn.
• Deliverable: Triages contacts into two queues—"verified for staging (Unsent)" and "ambiguities requiring human decision (FLAG)"—while synchronizing state across all systems.

The SDR Five-Step Closed Loop: From Data Fragments to Send Confirmation

In live prospecting, this system is tightly anchored to a standard SDR five-step progression:

Step 1: Research—Piecing Together Account Reality from True Product Footprints

Never break the ice with generic web search snippets or canned industry news. Cursor cloud agents query three authentic data sources:

1. Hex Data Footprints: Exactly how many engineers at the target company are actively trialing the product? Are active seats expanding or plateauing?
2. Salesforce CRM History: Has there been past interaction? Was the last closed-lost deal due to budget constraints, timing, or missing key product capabilities?
3. Gong Call Recordings: Pulls up past meeting transcripts to extract the prospect's verbatim business bottlenecks or specific tech stack decisions.

Step 2: Prospecting—The "Footprint Beats Title" Rule

Many outbound reps default to searching for VPs or Directors right away. This framework enforces strict tiering rules:

• Warm Relationships and Past Contacts First: If someone previously spoken to on Gong is still at the company, prioritize a Reconnect outreach;
• Cold Outreach Depends on Footprint Penetration: The system evaluates active product seats relative to software engineering headcount. When product footprint is strong, outreach targets Director-level and above (Director+) with matching responsibilities from a Golden List; if the footprint is nascent, targeting stays at Manager and above.

Step 3: Drafting—Leading with Verifiable Business Signals

Cursor strictly adheres to custom SDR Writing Skills, rejecting generic fluff like "Hope your week is off to a great start." The very first sentence must reference concrete account facts—such as: "Noticed your infrastructure team's telemetry volume on Hex has spiked recently..." or "Reviewing our Gong conversation from last November, when you were planning to migrate compute nodes to a new cluster...".

Step 4: System Intake—The Core CLEAR vs. FLAG Funnel

This is the standout architectural mechanism of the entire system. Every contact arriving from lead gen scrapers, Cursor morning research packets, or ChatGTM inboxes must undergo a rigorous cross-system audit by Grok Bot:

• CLEAR (Admitted to Staging): Verified that the contact is not actively enrolled in a Nooks sequence, not Paused in Amplemarket, and has no associated Stage 0 early-stage opportunity in Salesforce. Once cleared, the bot stages the draft into Nooks and assigns it to a sequence, with the status strictly marked as Unsent—serving purely as pre-staged holding rather than a zero-risk guarantee.
• FLAG (Held for Review): Intercepts ambiguous cases and holds them for human discretion. Contacts paused in Amplemarket, linked to Stage 0 opportunities in Salesforce, or showing duplicate or near-identical email addresses are routed to the human review backlog.

Step 5: Finalize—Human Review and Dispatch

When the SDR opens their laptop in the morning, they are greeted by a streamlined briefing:

• For items marked CLEAR, the human confirms the KEEP decision in Nooks and manually hits send;
• Blank LinkedIn connection requests (without custom notes) are dispatched separately by the dedicated bot (John Ford), representing an exception to manual execution;
• Reps focus their cognitive bandwidth on adjudicating the edge cases in the FLAG queue one by one.

The Specialized Bot Squad: Why "One Bot, One Job"?

In this architecture, Stefan did not build an all-knowing monolithic bot. Instead, he split responsibilities across five specialized agents with clear boundaries:

Boundaries of Autonomous Multi-Agent Collaboration

These bots collaborate directly with one another without requiring a human middleman: for example, once Josh Allen clears a batch of contacts, it instructs Contact Man to update the sheet and simultaneously notifies John Ford to dispatch LinkedIn connection requests within the same run.

Throughout this intake workflow, CLEAR contacts proceed to staging, while FLAG cases pause for human decisions; contacts with existing active touchpoints remain untouched to prevent duplicate outreach. CRM error remediation and sequence adjustments are handled separately according to each bot's dedicated role.

Sample Morning Triage: A 41-Contact Processing Run

The original author illustrated the workflow using a single morning's live batch (totaling 41 contacts as a breakdown of 6 + 34 + 1, representing an illustrative sample rather than an official aggregate or universal pass rate):

Morning Batch Processing Breakdown (Sample Run):

• Total Contacts in Sample: 41 contacts
• Already Active in Other Tools (Untouched / Skipped): 6 contacts
• Edge Cases Routed to FLAG for Human Review: 34 contacts
• Fully Cleared for Pre-Staging (CLEAR): 1 contact (reviewed and sent by author; blank LinkedIn request handled by John Ford)

This workflow routes distinct states to specialized handling paths:

1. Preventing Duplicate Outreach: Identifying 6 contacts already active in other tools, the bot leaves them untouched to avoid intrusive re-engagement;
2. Reserving Edge Cases for Human Judgment: The 34 flagged contacts represented boundary conditions—such as paused status in Amplemarket, association with early-stage opportunities in Salesforce, or duplicate/near-identical email addresses. The system aggregates these for human adjudication, eliminating manual cross-system sleuthing;
3. Batch Sequence Remapping: In one instance where 24 contacts were mistakenly placed in the wrong sequence, Josh Allen took over, closed out existing steps, remapped them to "email-only" or "call-only" tracks based on engineering titles, and scheduled initial calls for the following afternoon. Humans retained oversight of drafts and final approval, ensuring no follow-up fell through the cracks.

Five Rules for Preserving Pipeline Health

In standardizing this workflow into daily Standard Operating Procedures (SOP), the team established five non-negotiable rules:

1. Never auto-send Agents can research accounts, draft copy, create CRM objects, and stage sequences in Nooks, but the final "Send" button must be clicked by human sales reps. Human oversight is the last defense protecting brand equity against AI hallucinations.
2. Warm history beats a cold template If Gong records show even a single conversation from six months ago, the first-touch email must ground itself in verifiable specifics from that call. Never treat past contacts like strangers with boilerplate pitch templates.
3. Footprint beats title Active usage and seat depth on Hex dictate whether an account justifies SDR bandwidth to reach out to Directors or VPs. A polished executive title without any grassroots team adoption means jumping straight to leadership will only breed resistance.
4. LinkedIn requests stay blank The author configures LinkedIn connection requests without attached notes; the original text provides no comparative acceptance rate data for different note strategies. John Ford sends only plain connection requests; if a profile URL contains any ambiguity, the bot flags it rather than letting an LLM probabilistically "guess" if the profile is correct.
5. Repeated feedback becomes a rule When reviewing drafts, if a human rep catches the AI writing overly verbose copy twice in a row, or notices a certain job title shouldn't be enrolled in a phone track, they don't just edit that single email. They immediately codify that constraint into Cursor or Grok Bot persistent prompt and skill rule files.

Practical Assets: Ready-to-Adapt Prompts and Rule Templates

Adapting these patterns to your operational environment does not require an identical tech stack (for example, Hex corresponds to internal data warehouses or product analytics suites, Salesforce can be replaced by Fxiaoke or Xiaoshouyi, and Gong by Lark Minutes or DingTalk Flash Notes, and Nooks can be substituted with customized dialers or sequencing engines).

The core lies in the evaluation logic and boundary constraints within prompts. Below are two configuration examples adapted from the author's workflow (the original article did not publish the full prompt text; length constraints such as 4–5 sentences are editorial recommendations that can be adjusted; SKIP is an explanatory label introduced here for "leave untouched").

Asset 1: Cursor Research and First-Touch Drafting Specification (Markdown Format, Usable as Skill Rules)

# Role and Objective
You are a senior SDR research assistant supporting enterprise developer/technical products. Your mission is to synthesize the target account's product telemetry footprint, CRM history, and past call recordings into a structured account intelligence packet, and draft a high-converting first-touch outreach email for the SDR.

# Input Data Constraints
You must read and cite the following three factual sources; fabricating unverified data is strictly prohibited:
1. [Usage Footprint]: Analyze the target account's active seat count and adoption trends;
2. [CRM History]: Identify closed-lost reasons, churned contacts, or historical competitor evaluations;
3. [Call Archive]: Extract specific technical pain points, org changes, or blockers explicitly mentioned in past conversations.
# Outreach Tiering Rules (Footprint vs. Title)
- Compare active product seats against engineering team size (evaluated by [team-confirmed usage thresholds and org ratios]):
- If product footprint is nascent (below threshold):
  * Target Manager and above;
  * Director-level cold outreach requires additional verification of actual responsibilities and footprint criteria; never upgrade solely based on title.
- If product footprint is strong:
  * Select and target Director-level and above (Director+) with matching responsibilities from the Golden List.
  * Email angle: Focus on macro engineering efficiency and enterprise governance/security.

# First-Touch Email Drafting Rules
1. Zero opening fluff: Strictly prohibit phrases like "Hope you're having a great week," "Sorry to bother you," or "Reaching out because...";
2. First sentence must strictly be: A verifiable trigger point grounded in fact (such as an unresolved topic from Gong or a telemetry surge);
3. Keep the entire body within 4–5 sentences, in a professional, direct, restrained tone like an industry peer;
4. Call to Action (CTA): Propose a low-friction conversation (such as asking about feedback on a specific capability), without aggressively pushing a demo.

# Output Format
- [Account Diagnosis]: (Telemetry Overview / Historical Context / Icebreaker Trigger)
- [Recommended Targets]: (Name / Title / Rationales)
- [Email Draft]: (Subject Line + Body)

Asset 2: Grok Bot / Chief of Staff Intake and Triage Rules (CLEAR vs. FLAG)

# Role: Morning Prospecting Intake Reviewer (Chief of Staff Agent)
You are responsible for guarding outbound system health and brand reputation. Upon receiving contact lists from upstream, run rigorous cross-system checks and classify each contact as CLEAR or FLAG.

# Decision Flow
Evaluate every contact on the list against the following two rule sets:

1. [System Checks]:
   - Is the contact already active in Nooks? If already in active outreach, classify as [SKIP] (leave untouched, avoid duplicate outreach).
   - Is the contact Paused in Amplemarket? If paused, classify as [FLAG] (paused status must halt for human review; never classify as SKIP or automatically ignore).
   - Is the contact associated with a Stage 0 early-stage opportunity in Salesforce? If associated, classify as [FLAG] (defer to human sales judgment).
   - Are there suspicious edge cases, such as two nearly identical email addresses? If present, classify as [FLAG].

2. [LinkedIn Compliance]:
   - Check if today's dispatched connection requests have reached the Daily Cap;
   - Check the volume of Pending requests;
   - Verify the LinkedIn Profile URL. If UNVERIFIED, classify as [FLAG] (do not guess; intercept and hold for review).

# Execution Actions
- For [CLEAR]:
  * Signifies clearance for staging only; no zero-risk guarantee is implied;
  * Stage draft into Nooks sequence with status strictly set to "Unsent";
  * Contact Man updates sheet to log staged status, preventing duplicate handling;
  * John Ford verifies profile within quota limits and dispatches blank LinkedIn connection request.
- For [FLAG]:
  * Halt all automated downstream actions upon encountering any ambiguity or paused status; strictly prohibit auto-send, and route contact details to the SDR morning review queue.

Takeaways for Applying Agents to Real-World Work

For teams looking to operationalize agents in live enterprise environments, this SDR practice offers a grounded paradigm:

• Automation is not about replacing human decision-making, but elevating humans from manual chores to critical judgment: Sales reps are freed from tedious cross-system record juggling to focus on high-leverage decisions—such as determining the exact angle to re-engage an account with prior conversation history.
• Grant agents the right to stop: The hallmark of a production-grade agent system is not how fast it runs or how many messages it blasts, but whether it has the discernment to stop, raise a flag (FLAG), and ask for human review when encountering abnormal data, collision risks, or ambiguous profiles.

Letting models synthesize multi-source data and draft copy overnight, running background daemons for cross-system verification and staging, and leaving final dispatch and critical decisions to humans—this represents a pragmatic division of labor for human-agent collaboration in complex business workflows.

Template Links from the Original Article

• Add Josh Allen to Grok Bot
• Add Contact Man to Grok Bot
• Add John Ford to Grok Bot
• SFDC Create
• Sequencer
• GTM Account Researcher to Grok Bot

The links above are provided directly from the original article. After installation, you will still need to connect your own accounts, data, and tools; this review did not install or verify template states behind authentication.