Run a private personal-admin agent in Cursor
Series: Hub - One Agent, Many Workflows · Part 1 of 5 deep dives. Task: Series: Hub - One Agent, Many Workflows · Part 1 of 5 deep dives Why brainstorm and personal agent need different rules Brainstorming works best when the agent keeps options open. Personal admin works best when private details stay in a private lane. Both jobs can use the same agent interface. They should not share the same standing instructions. Who it is for: Anyone who uses AI for open-ended thinking and for private life logistics — and wants those two postures kept separate. What you will learn: how to separate explore mode from personal-agent mode with Customize, what to put in a handoff note, and a Path A you can run without Cursor. This post is a work-mode deep dive. For the full Customize map, start at the series hub. For harness framing, see lightweight harness in Cursor. For "is Cursor even for me," see Is Cursor only for developers?. When brainstorm and personal admin blur You open a chat to explore options for a paper, a product idea, or a career move. Twenty messages later you ask it to draft an email about a dentist appointment. The agent mixes the tone. The next day, a brainstorm thread may still carry a personal constraint that was never meant for a public document. The same blur shows up on always-on agents — hosted assistants like Hermes on my stack, or local agents like OpenClaw that keep one long memory thread across Telegram and chat apps. One personality, one memory store, many contexts. Without lane rules, work ideas, personal errands, and publishable drafts start sharing the same standing context. Cursor Customize (or equivalent rules in another harness) is how you mark which lane is active. Chat products encourage one continuous personality. Real weeks need two postures: Customize is how you teach the agent which posture is active. Principle: explore freely, commit in files The principle is simple. Exploration may be messy. Decisions that matter get written into a file you own. Personal facts stay in a personal lane with explicit privacy rules. You already do a version of this with separate notebooks or folders. The agent version is the same habit with standing rules so you do not restate "this is exploratory" every time. What "good" looks like in each mode Brainstorm mode Ask for options, risks, and questions - not a final plan on message one. Require the agent to label assumptions vs facts you provided. End with a short capture: keep / drop / park for later. Do not let the agent invent sources, quotes, or "I have seen teams do X" unless you supplied evidence. Personal agent mode Keep a small set of standing preferences (timezone habits, writing voice for private notes, what not to put in shareable docs). Prefer checklists and next actions over long essays. Never promote private details into blog drafts, proposals, or shared folders without an explicit ask. Use a dedicated handoff note so mobile captures do not vanish. Example implementation - how I run it I keep brainstorming and personal admin in the same Cursor interface, but I switch constraints deliberately. For brainstorm sessions I start with a one-line mode lock: explore, list unknowns, no fake decisions. I ask the agent to write keep/drop/park into a short ideas note when we stop. I review that note later; I do not treat chat scrollback as the archive. For personal-agent work I use a private notes area and rules that forbid copying personal details into Petralian drafts or commercial documents. On mobile I capture a follow-up in a few lines. At the desk I open that note and ask the agent to continue from Open questions. I direct the agent; I do not pretend I hand-wrote every checklist myself. Wiring detail for Brain + multi-project memory stays in the handbook. Here the point is the mode split. Tap diagram for full screen · pinch or ⌘/Ctrl+scroll to zoom when expanded Customize levers that matter most here You do not need every Customize layer for these two modes. If you later add consulting or blogging modes, their rules should be stricter. See Business development SSOT and Blogging and project memory. Mobile handoff for ideas and life admin On the move, brainstorming is often voice-of-thought: a risk you do not want to forget, a question for later, a title that might become a draft. Personal agent work is often a reminder: send this, buy that, prep for Thursday. Either way, close with a file update: One note per lane (ideas vs personal). Three headings: Context, Captures today, Open questions.; tools: Cursor. Human role: The original post describes the human role this way: I review that note later; I do not treat chat scrollback as the archive.; limitations: Mode locks require consistent human discipline; ignoring them renders Customize ineffective. Over-automating personal data poses privacy risks. Brainstorm notes are not research citations. The workflow depends on Cursor's Customize feature or equivalent rule-setting capabilities in other tools..
Tools
Published on Agent Work Atlas:
Overview
Series: Hub - One Agent, Many Workflows · Part 1 of 5 deep dives
Workflow
Series: Hub - One Agent, Many Workflows · Part 1 of 5 deep dives Why brainstorm and personal agent need different rules Brainstorming works best when the agent keeps options open. Personal admin works best when private details stay in a private lane. Both jobs can use the same agent interface. They should not share the same standing instructions. Who it is for: Anyone who uses AI for open-ended thinking and for private life logistics — and wants those two postures kept separate. What you will learn: how to separate explore mode from personal-agent mode with Customize, what to put in a handoff note, and a Path A you can run without Cursor. This post is a work-mode deep dive. For the full Customize map, start at the series hub. For harness framing, see lightweight harness in Cursor. For "is Cursor even for me," see Is Cursor only for developers?. When brainstorm and personal admin blur You open a chat to explore options for a paper, a product idea, or a career move. Twenty messages later you ask it to draft an email about a dentist appointment. The agent mixes the tone. The next day, a brainstorm thread may still carry a personal constraint that was never meant for a public document. The same blur shows up on always-on agents — hosted assistants like Hermes on my stack, or local agents like OpenClaw that keep one long memory thread across Telegram and chat apps. One personality, one memory store, many contexts. Without lane rules, work ideas, personal errands, and publishable drafts start sharing the same standing context. Cursor Customize (or equivalent rules in another harness) is how you mark which lane is active. Chat products encourage one continuous personality. Real weeks need two postures: Customize is how you teach the agent which posture is active. Principle: explore freely, commit in files The principle is simple. Exploration may be messy. Decisions that matter get written into a file you own. Personal facts stay in a personal lane with explicit privacy rules. You already do a version of this with separate notebooks or folders. The agent version is the same habit with standing rules so you do not restate "this is exploratory" every time. What "good" looks like in each mode Brainstorm mode Ask for options, risks, and questions - not a final plan on message one. Require the agent to label assumptions vs facts you provided. End with a short capture: keep / drop / park for later. Do not let the agent invent sources, quotes, or "I have seen teams do X" unless you supplied evidence. Personal agent mode Keep a small set of standing preferences (timezone habits, writing voice for private notes, what not to put in shareable docs). Prefer checklists and next actions over long essays. Never promote private details into blog drafts, proposals, or shared folders without an explicit ask. Use a dedicated handoff note so mobile captures do not vanish. Example implementation - how I run it I keep brainstorming and personal admin in the same Cursor interface, but I switch constraints deliberately. For brainstorm sessions I start with a one-line mode lock: explore, list unknowns, no fake decisions. I ask the agent to write keep/drop/park into a short ideas note when we stop. I review that note later; I do not treat chat scrollback as the archive. For personal-agent work I use a private notes area and rules that forbid copying personal details into Petralian drafts or commercial documents. On mobile I capture a follow-up in a few lines. At the desk I open that note and ask the agent to continue from Open questions. I direct the agent; I do not pretend I hand-wrote every checklist myself. Wiring detail for Brain + multi-project memory stays in the handbook. Here the point is the mode split. Tap diagram for full screen · pinch or ⌘/Ctrl+scroll to zoom when expanded Customize levers that matter most here You do not need every Customize layer for these two modes. If you later add consulting or blogging modes, their rules should be stricter. See Business development SSOT and Blogging and project memory. Mobile handoff for ideas and life admin On the move, brainstorming is often voice-of-thought: a risk you do not want to forget, a question for later, a title that might become a draft. Personal agent work is often a reminder: send this, buy that, prep for Thursday. Either way, close with a file update: One note per lane (ideas vs personal). Three headings: Context, Captures today, Open questions.
Reported result
The source does not report measured outcomes or success metrics. It describes a workflow intended to prevent tone mixing and privacy leaks. A proposed acceptance check is to verify that private details never appear in public drafts and that brainstorm notes remain distinct from personal logs after one week of use.
Human role
The original post describes the human role this way: I review that note later; I do not treat chat scrollback as the archive.
Limitations
Mode locks require consistent human discipline; ignoring them renders Customize ineffective. Over-automating personal data poses privacy risks. Brainstorm notes are not research citations. The workflow depends on Cursor's Customize feature or equivalent rule-setting capabilities in other tools.
Try it
I want to set up separate brainstorming and personal admin workflows in Cursor. First, create two distinct rule sets in Customize. For 'Brainstorm Mode', instruct the agent to list options, risks, and unknowns, label assumptions vs facts, and end every session with a 'keep/drop/park' summary. For 'Personal Mode', instruct the agent to provide only next actions, maintain a private tone, and strictly forbid copying any personal details into files marked for public sharing or commercial use. I will provide [replaceable inputs: specific brainstorm topic] and [replaceable inputs: personal task]. Run a dry run with one sample brainstorm and one sample personal task. Show me the output and the generated handoff notes. Do not save or publish anything until I confirm the privacy rules were respected and the brainstorm output was properly labeled. If the agent mixes tones or leaks private info, stop and ask for clarification on the rule definitions.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.