If you’re using Claude Opus 4.6 to write cold emails, you’re underusing it.
The real advantage isn’t in the message.
It’s in the research.
- In knowing who actually owns the outcome.
- In spotting the signal that changes the angle.
- In filtering out accounts that will never buy.
This guide walks through the exact system we use to turn Opus 4.6 into a research engine for outbound (while humans stay in control of judgement and sending).
The result is simple:
Fewer emails.
Better buyers.
Cleaner pipeline.
So you reach the right person, with the right reason, at the right time.
01What this is
This guide shows how to use an LLM as a research and signal engine.
It is not a messaging bot.
It does not send emails.
It does not book meetings.
It does not replace rep judgement.
What it does:
→ Cleans lists so reps stop wasting time
→ Finds the most likely buyer roles (not just titles)
→ Extracts real signals for real personalisation
→ Builds angles and proof so outreach is grounded
→ Helps teams run Tier 1 account plays without guesswork
How to use this guide:
→ Set your Inputs once (ICP + past wins + list format)
→ Run workflows in order
→ Use the prompt chains exactly as written
→ Review outputs before any outreach happens
02The system in one view
You will run 6 workflows:
The rule across all workflows:
→ If the model is unsure, it must say UNCLEAR and ask for what’s missing
→ Do not let the model guess to sound confident
03Inputs (set these once)
3.1 ICP filter (must be specific)
You need an ICP that works like a filter.
But you must take all the time you need to fill this out perfectly.
Minimum ICP fields:
→ Industry
→ Company size range
→ Common systems/tools (if relevant)
→ The buyer role (who owns the problem)
→ Trigger events (what makes them act now)
→ Disqualifiers (hard “no” rules)
3.2 Past wins (minimum 5, ideally 10 to 20)
You need real examples of how deals were won.
Minimum fields to capture per win:
→ Segment
→ Buyer role (actual decision-maker)
→ Trigger
→ Pain
→ Objections
→ Proof used
→ What closed the deal
3.3 Your lead list (format)
Each row should have at least:
→ Company name
→ Website
→ LinkedIn URL (optional but helpful)
→ Any notes you already have (optional)
04Workflow 1: Lead list cleaning
Goal:
→ Remove obvious non-ICP
→ Flag uncertain ones so you don’t waste rep time
Input:
→ A batch of leads (20 to 50 at a time)
→ Your ICP filter
Output:
→ YES / NO / UNCLEAR
→ One-line reason
→ What info is missing (if UNCLEAR)
PROMPT CHAIN: Workflow 1
05Workflow 2: Buyer role mapping (decision-maker discovery)
Goal:
→ Identify the most likely buyer role(s) for this account
→ Avoid title traps and wrong-person outreach
Input:
→ Company context you have (site, job posts, LinkedIn snippets, notes)
→ Your offer summary
→ Past win patterns (at least a short summary)
Output:
→ Primary buyer role candidate(s)
→ Why they likely own it
→ What would confirm it
→ If unclear, questions to resolve it
PROMPT CHAIN: Workflow 2
06Workflow 3: Signal extraction (real personalisation)
Goal:
→ Extract signals that actually change what you say
→ Avoid generic personalisation
Input:
→ Company snippets you have
→ The target buyer role from Workflow 2
Output:
→ 3 to 5 signals
→ Why each matters to that role
→ If you cannot find 2 strong signals, say INSUFFICIENT and list what to find
PROMPT CHAIN: Workflow 3
07Workflow 4: Angle builder (new value)
Goal:
→ Turn signals into specific angles a rep can use
→ Keep it grounded, role-specific, not generic value claims
Input:
→ Signals from Workflow 3
→ Your offer summary
Output:
→ 2 angles per signal
→ 1 risk of doing nothing
→ 1 upside if fixed
PROMPT CHAIN: Workflow 4
08Workflow 5: Proof builder (new value)
Goal:
→ Make angles believable with proof options
→ Prevent over-claiming
Input:
→ Angles from Workflow 4
→ Your proof assets (wins, results, systems, artefacts)
Output:
→ Proof options that match each angle
→ What is safe to claim vs not safe
PROMPT CHAIN: Workflow 5
09Workflow 6: Tier 1 account play (multi-contact plan)
Goal:
→ Build a simple multi-contact plan without spamming
→ Decide who to reach, in what order, and why
Input:
→ Buyer role
→ Signals
→ Angles + proof
Output:
→ 5 to 8 role types to include
→ Sequence order
→ Stop conditions
PROMPT CHAIN: Workflow 6
10Optional Workflow 7: Meeting quality gate (reply handling)
Goal:
→ Stop low-quality meetings
→ Redirect non-buyers fast
→ Keep conversations with the right person
PROMPT CHAIN: Workflow 7
11Metrics (keep it simple)
Track:
→ % of accounts that are ICP YES after cleaning
→ % of outreach going to true buyer roles (not title guesses)
→ % of replies from buyers (not juniors)
→ Meeting show rate
→ Pipeline created per rep per week (or per 100 accounts)
Do not optimise:
→ Raw reply rate
→ Volume for volume’s sake
12Common failure modes (quick list)
→ Too many UNCLEAR results
→ Your inputs are weak, not the model
→ Add better snippets (job posts, site copy, CRM notes)
→ Signals feel generic
→ You are feeding generic context
→ Add specific evidence (job post lines, product pages, role responsibilities)
→ Meetings with non-buyers
→ You skipped Workflow 2 or ignored UNCLEAR
→ Use the Meeting Quality Gate
13In conclusion
Claude Opus 4.6 is here to remove research waste and improve targeting.
Humans still decide who to pursue, what to say, and when to stop.
Run the workflows in order.
Keep outputs grounded.
If unclear, do not force it.
Questions people ask
Are you using AI agents to do this outreach?
In this system the model does the research and people do the outreach. It cleans lists, maps buyers, finds signals and drafts angles and proof. It does not send emails or book meetings, and a person reviews every output before anything goes out.
How do you figure out who to contact at a target account?
Buyer role mapping names the role most likely to own the outcome, with a confidence score. Below 70, the model asks three questions instead of guessing. For Tier 1 accounts, the account play then maps five to eight role types and the order to reach them in.
Software, accelerator and startup background. Runs the GTM architecture side: how the system is designed, what it is built on, and where AI actually earns its place.
LinkedIn ↗YouTube ↗