Automated Account Research with Claygents

Define the research questions before the prompt

The most common failure is a vague prompt like research this company. Decide the exact questions whose answers change your message. For most B2B motions, five:

• What does the company actually sell, in one sentence?
• What changed in the last 90 days (funding, launches, leadership, expansion)?
• What is the most likely pain relevant to your product, with evidence?
• Who owns that problem (function, not name)?
• Is there a concrete hook: a hiring post, a public statement, a tech signal?

Write the Claygent prompt with structured outputs

Add a Claygent column, choose a token-variable model for quality (Claude Sonnet or GPT-5.1), and enable structured outputs so each answer lands in its own column.

You are a B2B account researcher for a company that sells {{Your Product,
one line}} to {{Your ICP}}. Research {{Company Name}} ({{Domain}}).

Consult the company website, recent news, job postings and LinkedIn.
Answer ONLY from evidence you find; if you cannot verify something,
say "not found" rather than guessing.

Return JSON:
{"what_they_do": "",
 "recent_change": "",
 "likely_pain": "",
 "problem_owner": "",
 "hook": "",
 "confidence": 0.0-1.0,
 "sources": ["", ""]}

Test on 20 accounts you know well before scaling. In Sculptor, iterate the prompt conversationally and keep versions so you can roll back.

Draft the angle in a second column

Keep research and writing separate; it makes both debuggable. Add an AI column that turns the research into an opening angle, not a full email:

Using this research: {{what_they_do}} | {{recent_change}} | {{likely_pain}}
| {{hook}}

Write a 2-sentence opening angle for a cold email from us ({{Your Product}}).
Sentence 1 references the hook or change specifically and factually.
Sentence 2 bridges to the pain we solve. No flattery, no "I hope this finds
you well", no exclamation points. Write like a sharp colleague, not a marketer.

Reps get an angle to edit, which beats both a blank page and a fully ghost-written email they will not trust.

Gate on confidence and route to review

Set the drafting column to Only run if confidence is 0.7 or higher; below that, the account routes to a manual research queue rather than generating a weak angle from thin evidence. Create the review view where reps approve, edit or reject each row, and post a Slack digest when a batch of 25 is ready.

Push approved rows to the sequencer

Use the Outreach or Salesloft integration to add approved contacts to the right sequence with custom fields: {{custom_angle}}, {{research_summary}}, {{hook}}. The sequence template references the variables so every send is personalized without per-send writing. Write the research summary back to the CRM account so AEs and CS inherit it.

Close the loop with reply data

Sync reply and meeting outcomes back from the sequencer. Tag which hooks and pains earned replies, and feed the top patterns back into the prompt monthly. This is the feedback loop that separates compounding systems from decaying ones.