AI prompts for cleaning a messy CRM

You can use the prompts two ways. Paste a CSV export into Claude or ChatGPT and run a prompt on the file, then import the cleaned columns back. Or connect the AI client to your CRM over MCP, and the prompt works on live records with a full audit trail.

With SalesCrew, the second route takes one MCP token. Claude, ChatGPT, Cursor and custom agents all connect the same way, and every tool call is logged with the token that made it. See how to connect Claude to your CRM for the setup.

Find duplicate contacts in my CRM.
Match on: same email; or same first name + last name + company domain.
For each pair, show both records side by side: email, company, owner,
last activity date, number of open deals.
Recommend which record survives: the one with open deals first, then the
most recent activity. Do not merge anything. Wait for my list of approved pairs.

In SalesCrew this maps to duplicates_list for the pairs and duplicates_merge for the pairs you approve. Contact merges sit in the review class of the default agent policy, so an agent can propose them but a person confirms. The guide to merging duplicate records covers which fields survive.

Normalise company names on my contacts.
Rules: strip legal suffixes (Inc, LLC, Ltd, GmbH, Pty Ltd); keep the brand's
own capitalisation; use the website domain to confirm two spellings are one company.
Output a table: current value | proposed value | domain | confidence (high/low).
Only propose changes where confidence is high. Change nothing yet.

Company names drift fastest because every import spells them differently. Confirming on the domain stops the model from merging two real companies that share a word.

Fill missing job titles.
Find contacts with an empty job title. For each, read the last three email
signatures in the thread history and extract the title as written.
If no signature holds a title, leave the field empty.
Never overwrite a title that already exists. Show me the list before saving.

Standardise location fields.
Convert country to the full English name (United States, United Kingdom).
Convert US states to two-letter codes. Leave city as written.
List every value you could not map so I can fix it by hand.

Clean location values matter the day you build a segment by country or send inside a business-hours window. It's a boring fix that pays off every week after.

Flag role-based and personal addresses.
Mark contacts whose email starts with info@, sales@, admin@, support@,
hello@, office@ or contact@ as role addresses.
Mark gmail.com, outlook.com, yahoo.com and icloud.com addresses as personal.
Return counts per group and the full list. Do not delete or suppress any.

Find deals that are dead but still open.
List open deals with no email, meeting, task or note in the last 45 days.
For each, give: deal name, stage, value, owner, last activity and what it was.
Suggest one of: close as lost (with a reason), set a follow-up task, or keep.
Do not change any stage. I will approve each suggestion.

SalesCrew's stalled_deals tool returns the open deals that have gone quiet. Stage changes stay in review, so the forecast only moves when you approve the change.

Rewrite messy notes into a clean summary.
For the contact I name, read every note and email thread.
Write a five-line summary: who they are, how they found us, what they asked
for, where the deal stands, the next step with a date.
Quote the contact's words where they made a commitment. Save it as a new note
only after I confirm.

Audit required fields on active deals.
For every open deal, check: amount, close date, stage, owner, primary contact.
Return a table of deals missing any of these, with the missing fields named.
Sort by deal value, largest first.

Running cleanup once a quarter leaves three months of drift. Put prompts 1, 3 and 6 into a Claude Skill and run it every Friday. The Claude Skills for sales page has a ready crm-hygiene skill built from these prompts, and the CRM implementation plan shows where cleanup fits before a new CRM goes live.