KoltrixPrompting an AI assistant to search your mail effectively
How AI assistants turn your requests into mailbox searches, and how to phrase prompts with names, dates and domains to find the right thread the first time.
On this page(10 sections)
- What happens when you ask
- Give the assistant search handles
- Names and addresses
- Dates and time ranges
- Distinctive words
- Which mailbox or label
- Vague vs precise: worked examples
- Ask for links, not just answers
- Ask for the candidates when unsure
- Verify before you act
- Prompts worth saving
- When search isn't the right tool
- A note on access
- Bottom line
Asking an AI assistant "find that email about the contract" feels like it should work. Sometimes it does. Often you get the wrong contract, three irrelevant threads, or a confident summary of an email that isn't the one you meant.
The reason is simple: behind the conversation, the assistant is still running ordinary mailbox searches. Knowing how it translates your request into those searches lets you write prompts that land on the right thread the first time.
What happens when you ask
When an assistant is connected to your mailbox (through MCP or another integration), it can't scan every email you've ever received. It calls a search tool with a query, gets back a short list of matching threads, and then reads some of them.
So a request like "What did the designer say about the logo?" becomes something like:
- The model picks search terms: maybe
logo designer. - The mailbox returns threads matching those words.
- The model reads the top few results.
- It answers based on what it read.
Every weak point is visible here. If the search terms are vague, the results are noisy. If the right thread isn't in the top results, the model never sees it. And if it reads the wrong thread, it may answer confidently anyway.
Give the assistant search handles
The fix is to hand the assistant the details a search engine can use.
Names and addresses
People's names, company names and email domains are the most useful search terms. "The designer" means nothing to a search index. "Lena from Brightline" or "anyone at brightline.example" does.
Dates and time ranges
"Last week", "in March", or "before the launch" narrow results dramatically. Assistants can usually translate relative dates into a date filter.
Distinctive words
Think about words that were actually in the email: an invoice number, a product name, a project codename, an unusual phrase. Generic words like "update" or "question" match everything.
Which mailbox or label
If you have access to several mailboxes, say which one: "in support@" or "in my own inbox". If you use labels, mention them: "threads labeled billing".
Vague vs precise: worked examples
| Vague prompt | What tends to happen | Precise prompt |
|---|---|---|
| "Find the contract email." | Many threads mention contracts; the assistant picks one, maybe the wrong one. | "Find the thread with Omar at Fieldstone about the renewal contract, from September." |
| "What's the status of that bug?" | No idea which bug; it may summarize the newest bug-related email. | "Search support@ for threads about CSV export failing in the last two weeks. List each with the customer and latest status." |
| "Did the accountant reply?" | Searches for "accountant", which may not appear in the email. | "Did anyone from ledgerworks.example reply since Monday?" |
| "Show me the important emails." | "Important" isn't searchable; results are arbitrary. | "List unread threads from customers in Primary from the last three days." |
The precise versions aren't longer for the sake of it. Each added detail is a filter the assistant can apply.
Ask for links, not just answers
When the assistant finds something, ask it to show you where:
Find the thread with Omar about the renewal contract and give me the subject, date of the latest message, and a link to the thread.
This does two things. You can open the thread and check the assistant's reading. And if the assistant can't produce a specific thread, that's a signal it's guessing.
Ask for the candidates when unsure
If you're not certain which thread you mean, ask for options rather than an answer:
Search for threads mentioning "onboarding call" from the last month. List up to five with sender, subject, and date. Don't summarize yet.
Then pick the right one and ask your actual question about it. Two short steps are more reliable than one ambiguous step.
Verify before you act
Assistants occasionally merge two threads in their answer, mix up who said what, or present an older message as the latest. Before acting on anything important:
- Open the thread using the link.
- Check the latest message yourself, especially for dates, amounts, and decisions.
- Be suspicious of tidy answers to messy questions. If you asked about a dispute and got a neat resolution, read the source.
This matters most when you're about to reply, commit to something, or forward information.
Prompts worth saving
A few patterns that work well and are worth keeping in a note:
Find threads from [person or domain] about [topic] since [date].
List subject, date of latest message, and link. Don't summarize yet.
In [mailbox], list unread threads from customers in the last [N] days
that are waiting on a reply from us. One line each with a link.
Summarize the thread [link or subject] in five bullets: status,
decisions, open questions, who owes what, key numbers (copied exactly).
Search for emails containing invoice number [number]. Who sent it,
when, and was there any reply?
When search isn't the right tool
Some questions are hard for any search-based assistant:
- "What haven't I replied to?" depends on the assistant being able to tell reply state, which varies by integration.
- "Everything about this customer, ever" can return more threads than the assistant reads; it may summarize only a sample.
- Attachments' contents may not be searchable or readable depending on the tool.
For these, combine the assistant with your email app's own views and filters rather than relying on one prompt.
A note on access
An assistant can only search what it's connected to. Koltrix's MCP server gives assistants like Claude and ChatGPT read access to your mail, plus the ability to organize and draft, but never to send.
Bottom line
- Behind the chat, the assistant runs ordinary searches, so give it searchable details.
- Use names, domains, date ranges, distinctive words, and the mailbox or label.
- Ask for thread links and candidate lists, not just answers.
- Verify the source thread before acting on anything that matters.
- Save the prompts that work; precise patterns get reused constantly.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.
