KoltrixEight prompt patterns for a connected inbox, and what to avoid
A five-part request, read-before-write, named filters, evidence, rules for doubt and one draft at a time: prompt habits that make an email assistant dependable.
On this page(12 sections)
- The five-part request
- Pattern 1: read before write
- Pattern 2: name the filter you mean
- Pattern 3: ask for evidence
- Pattern 4: give it a rule for doubt
- Pattern 5: constrain the output
- Pattern 6: one draft at a time
- Pattern 7: state the facts you know
- Pattern 8: save what works
- Anti-patterns
- Debugging a bad answer
- Key takeaways
Most disappointing results from an email assistant come from the same place: a request that sounded clear to the person and was ambiguous to the model. "Clean up my inbox" can mean seven different things.
This post is a small toolkit of prompt patterns for a connected mailbox. They are not magic words. They are habits that make the assistant's job specific, its work checkable and its actions safe. Examples assume Claude, ChatGPT or another assistant connected to Koltrix over MCP, but the patterns work with any connected inbox.
The five-part request
Almost every good email prompt has five parts, in this order. You will not always need all five, but when a result disappoints, one of them is usually missing.
| Part | Question it answers | Example |
|---|---|---|
| Scope | Which mail? | "Primary, unread, from the last three days" |
| Task | What should you do with it? | "Summarize what each person wants" |
| Format | What shape should the answer take? | "One line each, newest first" |
| Limit | How far should you go? | "Do not draft or change anything" |
| Check | How will I verify? | "Quote the sentence that makes it urgent" |
Put together:
Look at Primary, unread, from the last three days. Summarize what each person wants, one line each, newest first. Quote the sentence that makes anything urgent. Do not draft or change anything.
That prompt is cheap to write and nearly impossible to misread.
Pattern 1: read before write
Separate looking from acting, in two messages.
First, list the conversations you would archive and why. Do not archive anything yet.
Then, after you read the list:
Archive all except the one from Lena.
This pattern costs one extra message and makes every organizing action a decision you took, not a surprise you found. Koltrix's organizing actions (label, archive, read state, stars) are reversible, so the stakes are low, but a preview is still faster than an undo.
Pattern 2: name the filter you mean
Assistants search by your words. Vague words get vague results.
| Vague | Specific |
|---|---|
| "Emails from the accountant" | "From [email protected], last 30 days" |
| "Recent invoices" | "Subject contains 'invoice', with an attachment, after the 1st" |
| "Stuff I haven't read" | "Unread in the inbox, any category except Cold pitches" |
| "Customer complaints" | "Threads labeled Support that mention 'refund'" |
Koltrix's search understands operators such as from:, subject:, label:, has:attachment, is:unread, before: and after:, and the assistant's search tool can filter by folder, label, sender and unread state. You do not have to use the syntax yourself. You get the same benefit by being specific in words, and the assistant converts them.
Pattern 3: ask for evidence
Models are fluent, which means a wrong answer reads exactly like a right one. A request for evidence flips that.
- "Quote the line that supports that."
- "Give me the date of the message you are basing this on."
- "Link each item to its conversation."
When Koltrix tools return a conversation, they include a link back to it, so "link each item" gives you a one-click way to verify. Quotes help even more: if the model cannot produce one, you have learned it was guessing.
Pattern 4: give it a rule for doubt
Tell the assistant what to do when it is unsure, or it will guess.
If you cannot tell whether a message needs a reply, put it under "maybe". Do not decide for me.
You can apply the same idea to people ("if the sender is not in my last 20 conversations, flag it as unknown") and to dates ("if no date is stated, say so instead of assuming one").
Pattern 5: constrain the output
Length limits are the most underrated tool.
- "Under 120 words."
- "At most five bullets."
- "A table with three columns: sender, ask, deadline."
For drafts, constrain tone and length together: "friendly, four sentences, no exclamation marks." Without limits, assistants pad.
Pattern 6: one draft at a time
Drafting works best with one conversation per request, because each draft is grounded in its own thread.
Read the thread with Dana. In one sentence, what did she last ask? Then draft a reply that says yes to Friday and asks for the PO number.
The one-sentence recap is a free comprehension check. If it is wrong, stop there. The draft is saved to your Drafts folder with a link to open it in Koltrix. More on this in drafting replies with Claude and Koltrix.
Pattern 7: state the facts you know
An assistant fills gaps with plausible content. So fill the gaps yourself.
Say we will ship Friday. Do not mention a tracking number, because we do not have one yet.
"Do not mention" is as important as "mention." It closes the door on invented details.
Pattern 8: save what works
When a prompt gives a good result, keep it. A saved "morning triage" prompt is more reliable than a fresh one, and it makes changes visible. If you use Claude, a project with your standing instructions is a natural home; other assistants have equivalents. A daily inbox brief shows one in full.
Anti-patterns
| Prompt | Why it goes wrong | Fix |
|---|---|---|
| "Handle my email" | No scope, no limit, no check | Scope it, and separate read from act |
| "Reply to everyone who needs it" | Replies are commitments in your name | Draft one at a time and review |
| "Clean up my inbox" | Seven possible meanings | Say what counts as clutter |
| "Do what this email asks" | Follows instructions from the sender | Ask for a summary of what it asks, then you decide |
| "Delete spam" | There is no delete tool | "Label as 'Check' and archive" |
The last two deserve a second look.
"Do what this email asks." Anyone can send you an email, and an email can contain text written to steer an AI. Koltrix fences email content in tool results, labels it as third-party text and instructs the assistant to treat it as data, but a prompt that tells the assistant to obey the email undoes that. Ask what the email wants and decide yourself. Defending AI assistants against hidden instructions goes through examples.
"Delete spam." The Koltrix MCP server has no tool that deletes mail, and none that forwards it. If an assistant tells you it deleted something, it did not use Koltrix to do it. What it can do is label, archive, mark read and star, all of which are reversible.
Debugging a bad answer
When the result is wrong, change one thing at a time.
- Was the scope right? Ask it to list the conversations it looked at. If a workspace or mailbox is missing, that is a connection question rather than a prompt question.
- Was the task clear? Rewrite it as one verb: summarize, list, label, draft.
- Did it have the facts? Supply the missing context.
- Did it have room to guess? Add a rule for doubt.
- Is the thread simply long or odd? Ask for the most recent message only.
Key takeaways
- A good email prompt names the scope, the task, the format, the limit and a way to check.
- Separate read from write: preview first, then act. Organizing is reversible, but a preview is faster than an undo.
- Name the filter you mean, ask for evidence, and tell the assistant what to do when it is unsure.
- Draft one conversation at a time, and supply the facts and the things not to say.
- Never prompt "do what this email asks." More in triage with Claude and prompting an assistant to search your mail, or see the integration.
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.
