How to give Claude precise instructions when using connectors
In brief
Vague instructions get vague results. The difference between "find me that doc in Notion" and an instruction that actually works — every time.
Contents
There is a version of working with connectors that frustrates people, and a version that feels almost automatic. The difference is not the connector itself. It is how precise the instruction is.
When you give Claude a vague instruction — "find me that Notion doc about the Q3 roadmap" — you are asking it to interpret what you meant, search in a way that matches your mental model of how the data is organized, and return something useful. Sometimes it works. Often it does not, and you end up re-asking or doing the search yourself.
The more precise instruction removes interpretation from the equation. Claude does not have to guess at what you mean. It knows exactly what to look for, where, and what to return.
This is a guide to writing connector instructions that work reliably — and the mental model behind why precision matters.
What a connector actually does
A connector gives Claude access to an external tool — Notion, Google Drive, Slack, a CRM, a calendar, a database. When you ask Claude to do something that requires information from that tool, it queries the tool using specific parameters and brings back the results.
The quality of what comes back depends almost entirely on how specific those parameters are. A search query like "roadmap" returns everything that contains the word roadmap. A search query for "pages in the Product workspace tagged Q3 modified after July 1" returns something far more useful.
You are not just describing what you want to a human who will use judgment to fill the gaps. You are specifying a query. The more specific the query, the more useful the result.
The four elements of a precise instruction
Every effective connector instruction has four components. Not all are needed every time, but knowing them helps you figure out what is missing when an instruction fails.
1. The resource type
What kind of thing are you looking for? A document, a row in a database, a message, a task, an event, a contact? Being specific about the resource type focuses the search dramatically.
Vague: "Find the roadmap in Notion."
Precise: "Find the Notion page titled Q3 Product Roadmap."
If you do not know the exact title, narrow by type and location: "Find Notion pages in the Product workspace that contain 'Q3 roadmap' in the title."
2. The location or scope
Where should Claude look? Most tools have structure — workspaces, folders, channels, projects, databases. Telling Claude where to look prevents it from searching everything and returning a confusing mix.
Vague: "Check Slack for the conversation about the design feedback."
Precise: "Search the #product channel in Slack for messages about design feedback posted in the last 14 days."
3. The filter or qualifier
What properties should narrow the results? Dates, authors, tags, statuses, assignees, labels — most tools have metadata you can filter on. Using it turns a broad search into a targeted one.
Vague: "Pull the open tasks in Linear."
Precise: "Pull Linear issues assigned to me with status In Progress or In Review, sorted by priority."
4. What to return
What should Claude do with the results? Summarize them, list them, extract specific fields, draft a response based on them? Specifying the output format prevents Claude from returning a raw dump when you wanted a clean summary.
Vague: "Get my calendar for tomorrow."
Precise: "Get my calendar events for tomorrow between 8am and 6pm. List each one with the title, time, and any attached meeting link. If there are back-to-back meetings with no break, flag them."
The before/after pattern
Here is what this looks like in practice across common connectors.
Notion
Before: "Find the notes from our last team meeting."
After: "Find the Notion page in the Team Meetings database with the most recent date. Return the title, date, and the section called Action Items."
Google Drive
Before: "Find the proposal we sent to Acme."
After: "Search Google Drive for files with 'Acme' and 'proposal' in the title, modified in the last 60 days. Return the file name, last modified date, and the link."
Slack
Before: "See what people said about the launch."
After: "Search the #announcements and #general channels for messages mentioning 'launch' posted between October 1 and October 15. Summarize the main reactions and any follow-up questions that were raised."
Linear / Jira
Before: "What are we working on?"
After: "Pull all Linear issues in the Current Sprint cycle assigned to the Product team. Group them by status: In Progress, In Review, Done. Flag anything with a due date in the next three days."
CRM (Salesforce, HubSpot)
Before: "Check the status of the Acme deal."
After: "Find the Acme Corp account in HubSpot. Return the deal stage, last activity date, next follow-up task, and the name of the deal owner."
When to save instructions as defaults
If you find yourself writing the same precise instruction repeatedly — your daily task list, your weekly meeting notes, your inbox triage — save it as a default instruction in your system prompt or project context.
The instruction "every morning, pull my Linear issues in progress, my calendar for today, and any Slack DMs I haven't replied to — give me a single prioritized list" becomes a morning routine you run once, not an instruction you type from scratch every day.
The connectors do not need to be re-explained each time. They need to be set up once with precision, tested until they work reliably, and then triggered without friction.
The debugging pattern when it does not work
When a connector instruction returns the wrong thing, the problem is almost always in one of three places:
The search terms are too broad. The word you used appears in many more places than you expected. Add qualifiers — date ranges, specific fields, exact phrases in quotes.
The scope is wrong. Claude searched the right tool but the wrong part of it. Specify the workspace, folder, channel, or project explicitly.
The output is not specified. Claude returned the raw results instead of processing them. Add a final sentence describing what you want done with the results.
Run through these three checks before concluding the connector does not work. In most cases, the connector is fine — the instruction is just not specific enough.
Further reading
- How to write precise instructions — getting the most from connectors