KoltrixInbox triage with Claude: prompts that work and why
A three-pass method and a set of prompts for triaging a Koltrix inbox with Claude or ChatGPT, with notes on the tools underneath so you can predict the results.
On this page(8 sections)
Triage is the job where a connected assistant earns its keep fastest. You are not asking it to be creative. You are asking it to look at forty conversations and tell you which three matter.
This post is a set of prompts and a method for using Claude (or ChatGPT, or any MCP assistant) to triage a Koltrix inbox, with notes on what the tools underneath actually do, so you can predict what the assistant will and will not get right.
Know what it is working from
Koltrix already sorts incoming mail before any assistant gets involved. Every new message gets a category (Primary, Other or Cold pitches, among others in the app) and a one-line summary where it has one. The assistant can use that sorting instead of redoing it.
Three tools do most of the triage work:
| Tool | What it gives the assistant |
|---|---|
get_inbox_summary |
Unread counts for the inbox overall and per category, plus the conversations Koltrix thinks need a reply |
list_threads |
A page of conversations, newest first, narrowed by folder, category (primary, other or cold_pitch), label or unread |
get_thread |
Every message in one conversation as plain text, with attachments listed by name and size |
Two details are worth knowing. get_thread does not mark a conversation read, so asking for summaries will not quietly clear your unread count. And the list tools return at most 50 conversations per page, with a cursor for the next, so very large backlogs are handled in pages.
The three-pass method
Resist the urge to ask "triage my inbox" in one go. Three short passes work better, because each one gives you a checkpoint.
Pass 1: the map
How many unread conversations do I have in each category? Then list the ones Koltrix thinks need a reply, with sender and subject.
You get counts and a short list. This costs the assistant one or two tool calls, and you can compare it with the Koltrix sidebar to confirm it is looking at the right workspace.
Pass 2: the reading
For each of those conversations, tell me in one sentence what is being asked of me and whether there is a deadline. Do not draft anything yet.
"Do not draft anything yet" is doing real work. It keeps the assistant from running ahead, and it stops you from ending up with six drafts of the wrong things. Notice that the assistant has to read the threads to do this, and your job is to spot-check two of them against the originals.
Pass 3: the decision
Rank them by urgency. Put anything with a date in the next two days first. Then tell me which I could answer in under a minute.
Now you have a plan. You choose what to do; the assistant helped you see the shape of the inbox.
Prompts that work, and why
Scope by category.
Look only at Primary, unread, from the last three days.
The Primary category is where people writing to you land. Starting there avoids spending the assistant's attention on newsletters and cold pitches. The list_threads tool can filter on that category directly.
Ask for evidence.
For each conversation you flag as urgent, quote the sentence that makes it urgent.
When the assistant has to quote the thread, it either did read it or it fails visibly. Quotes also let you check fast.
Give it a rule for ambiguity.
If you are not sure whether something needs a reply, put it in a "maybe" list rather than guessing.
An explicit "maybe" bucket prevents confident misclassification.
Separate reading from acting.
Summarize first. Wait for me to say which ones to label or archive.
Organizing is reversible, but you still want to choose.
Name the people who matter.
Treat anything from acme.com or from Dana as high priority.
Assistants do not know your priorities until you say them. The search_mail tool can filter by a full sender address or a whole domain, so this kind of instruction maps cleanly onto real filters.
Acting on the result
Once you have a plan, you can have the assistant do the mechanical part. These are the organizing tools Koltrix provides: add or remove a label, archive, mark read or unread, and star or unstar. All of them can be undone in the app.
Star the three conversations you ranked highest. Label the "maybe" list "Check". Mark the newsletters read.
A reminder on labels: if you name a label that does not exist yet, Koltrix creates it. A typo creates a new label, so glance at the spelling or ask the assistant to list your labels first. Organizing your inbox by chat covers this in depth.
For the ones to answer, ask for drafts one at a time, which keeps each draft grounded in its own thread:
Draft a reply to the Acme thread: confirm Friday, ask for the PO number. Keep it to four sentences.
The draft lands in your Drafts folder with a link to open it in Koltrix. By default, sending is off, so nothing leaves until you click Send in the app.
What a ten-minute session looks like
Here is the shape of a realistic session. The times are a rough guide, not a measurement:
| Minutes | Step | Output |
|---|---|---|
| 0 to 1 | Pass 1: map | Counts per category and a list of threads waiting on you |
| 1 to 4 | Pass 2: reading | One-line asks and deadlines for each, you spot-check two |
| 4 to 5 | Pass 3: decision | A ranked list with the quick wins marked |
| 5 to 7 | Organize | Stars, labels, newsletters marked read |
| 7 to 10 | Draft | Two or three drafts waiting in Drafts for your edit |
You still read and send the replies. What changed is that you started the day knowing what mattered.
What to watch for
- Confident wrong answers. Summaries can drop a condition, such as "only if legal approves." Open a thread before acting on anything with consequences. The assistant gives you a map, not the territory.
- Truncation. Very long threads may be summarized unevenly. Ask "what is the most recent message asking for?" if the thread is long.
- Hidden instructions. Email bodies may contain text written to steer an AI. Koltrix fences email content in tool results and labels it as untrusted, third-party text, and the assistant is told to treat it as data. Even so, if a summary ever says something like "this email asks you to do X," read the email yourself. Defending AI assistants against hidden instructions is worth a read.
- Wrong workspace. A connection belongs to one workspace. If the counts look off, check which one you approved.
Make it repeatable
Once a prompt works, save it. A short "morning triage" prompt that you paste every day is more reliable than improvising, and it gives you a baseline to notice when something changes. A daily inbox brief with Claude and Koltrix turns it into a routine, and MCP email prompt patterns that work explains why these prompts are shaped the way they are.
Key takeaways
- Triage is the strongest early use of an assistant: it reads many conversations and tells you which matter.
- Use three passes (map, read, decide) instead of one big request, and say "do not draft yet."
- Lean on Koltrix's own sorting:
get_inbox_summaryand the category filter onlist_threadsgive the assistant a head start. - Ask for quotes, give rules for ambiguity, and spot-check before acting.
- Organizing is reversible and drafting is reviewable, which is why this is a comfortable place to start. See what the integration offers.
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.
