Manage your tasks from Claude and ChatGPT

Connect your AI assistant to your task manager through MCP. How to set it up, what to ask, real prompts to start with, and the habits that keep it safe.

6 min read

You can connect an AI assistant to your task manager, so it works on your real tasks, projects and notes instead of on text you paste into a chat. The standard that makes this possible is the Model Context Protocol (MCP), an open way for AI applications to connect to external tools and data. This guide covers why to do it, how to connect, what to ask, and how to keep control.

#The method in a few lines

1. Connect the assistant once, and approve it in your browser.
2. Start with questions that only read your data.
3. Ask for changes in small steps, and ask it to show you what it proposes before it writes.
4. Keep the confirmations on for anything that deletes or reaches other people.
5. Check what is connected from time to time, and remove what you no longer use.

#Why connect an assistant

The chat is where many people already think: they paste an email, a lecture or a meeting transcript and ask what to do with it. Without a connection, the last step is yours: copy the result, open the task manager, create the tasks one by one. With a connection, the assistant does that step.

• It sees your real list. "What should I do today?" gets an answer based on what is overdue and due, not on a guess.
• It does the bulk work. Creating forty tasks from a plan, triaging an inbox, or writing a deck of flashcards from a course takes one request.
• It stays where you work. You do not switch apps. What it changes appears in your Prior apps at once.
• It uses your own assistant. The work runs under your Claude or ChatGPT subscription, not Prior's. The connection itself works on every Prior plan.

It has limits. An assistant can misread a request or pick the wrong project. Prior never starts a conversation with it: it acts when you ask. And it has access to your real data, which is why the last part of this guide matters.

#How to connect

The first time, your browser opens Prior so you can sign in and allow the connection. You see which app is asking before you click Allow.

claude mcp add --transport http prior https://api.priortasks.com/mcp

#What to ask

Start by reading. Ask a question that cannot change anything, and see how your assistant answers.

Prior also offers four workflow prompts, plan_my_day, triage_inbox, weekly_review and break_down_project, and a fifth for flashcards, make_flashcards. Clients that list MCP prompts show them, and Claude Code shows the flashcard one as /mcp__prior__make_flashcards. Each workflow prompt asks the assistant to show what it proposes before changing anything.

If a task or a note holds a word you want the assistant to find, say so: it can search across tasks, notes, projects and habits. More recipes are in the MCP recipes.

#Habits that keep it safe

• Read first. Begin each new kind of request with a question that only reads.
• Ask for the plan before the write. "Show me what you propose, then apply only what I accept" is a good sentence to add to any request.
• Name the scope. "In the Website project" is a safer instruction than "in my tasks".
• Keep confirmations on for deletes and for sharing. Prior labels its tools: some only read, some change data, some delete, and a few reach people outside your account, such as inviting someone to a project or creating a share link. Clients can use those labels to decide when to ask you first. Deleting a note moves it to the trash and a deleted project can be restored, but there is no restore tool for tasks.
• Prefer small edits to notes. Ask for an insertion or a changed paragraph, not a rewrite of the whole note. Prior tells the assistant to use edits or appends on long notes.
• Treat other people's text as data. Comments, shared notes and task titles written by others could contain instructions aimed at an assistant. Prior tells the assistant to treat what it finds in tasks, notes and comments as data, not as instructions. That helps, but it does not make it safe to approve a change you did not read.
• Draft messages, then send them yourself. Ask the assistant to write a comment or an invitation, and read it before anything goes out.
• Mind the day. Prior's tools use UTC when no date is given. If you work late in another time zone, name today's date in your request.

The assistant has your permissions and no more. It sees your own data and the projects shared with you, and your role in a project still applies: a viewer cannot edit through an assistant either. Your password, sessions, access keys, subscription and payment details, and account deletion are out of its reach.

#Mistakes to avoid

• Starting with a big write. "Reorganize everything" is a request you cannot check. Start with one project.
• Approving without reading. The review is the whole point. Read the list of changes.
• Pasting untrusted text and saying "do what it says". Text from a web page or an unknown sender can contain instructions of its own.
• Leaving old connections. Remove the apps you no longer use.
• Expecting memory between chats. A new conversation may know nothing about the last one. Name the project again.

#In Prior

• Settings → Integrations holds the install buttons, a command for Claude Code and the list of Connected apps. Revoke disconnects one. Clients without browser sign-in can use an access key.
• A task completed through an assistant earns XP like any other, and Game mode shows it when you come back.
• Read Use Prior from your AI assistant for the overview, then the pages for Claude, ChatGPT, Claude Code and Cursor and VS Code.
• The tools lists everything your assistant can reach, and how it stays safe explains permissions and sign-in. The public MCP page lists the clients Prior supports.