PersonalMaestro:SupervisedTaskAutomation

Every home and personal project board has that one column where tasks go to die.

You need to look up municipal setback requirements for a new fence, find three licensed heat pump installers in your county, or compare trailer hitch tongue weights. None of these tasks are mathematically difficult. But they take forty-five minutes of staring at terrible government websites, cross-referencing PDFs, and compiling phone numbers.

Tasks that, for lack of a better word, just suck.

Because you have a day job, kids’ soccer games, and real-world responsibilities, those cards sit in your backlog for three months. By the time you finally sit down with a double espresso on a Saturday morning to knock them out, the weekend is half over.

I built Personal Maestro to turn those stagnant backlog cards into finished briefing packets while I’m away from the computer.

You drag a card into Doing. Maestro reads the card, compiles a clean briefing pack, assigns it to a specialist worker, and stages the research directly on the back of the card. When you open your board, the research is done, the contractor email is drafted, and all that’s left is for you to make the call.

Inspect the live task queue, execution status, and human approval boundary in the interactive rig below:

Why the Architecture Matters

Letting an autonomous script touch your personal boards can quickly create a mess if you don’t build hard guardrails into the code. I designed the system around three non-negotiables:

1. The Column Is the Assignment

Nothing enters Maestro Doing automatically. The background supervisor doesn’t scan the entire board or invent its own initiatives. Moving a physical card into the column is the deliberate human gesture that triggers execution.

2. Drafts Never Need Permission. Sends Always Do.

An agent can research local codes, summarize vendor pricing, and draft an introductory email to a fencing contractor. But it physically lacks the API tool to click “Send.” Outbound correspondence stays parked in Needs Review with a bright orange badge until a human reviews the text and confirms it.

3. The Fire Emoji Gate

When an agent stages an action that touches money or live communication, it posts a crisp summary to the console: recipient, cost, and exact message body. A generic “yes” or “looks good” does nothing. The only token that releases the lock is the 🔥 emoji. The weirdness is the feature: you cannot type it by accident.

Specialist Worker Roles

Instead of giving one giant agent unrestricted access to everything, tasks are handed off to isolated specialist scripts:

• pm-researcher: Pulls county permit requirements, compares hardware dimensions, and formats product comparison tables into local Obsidian notes.
• pm-planner: Takes a messy goal (like “install a 50-amp RV hookup”) and breaks it down into sequential, prerequisite steps with milestone estimates.
• pm-writer: Drafts clear, concise contractor inquiries and follow-up emails, staged in draft mode with zero external write permissions.

Honest Caveats

• Hallucinated Municipal Details: Local government websites are notorious for hosting conflicting PDFs from 2014 and 2022 on the same domain. The researcher node will occasionally cite an outdated permit fee. Verification of key figures against the actual building department phone line remains a human job.
• Process Discipline Required: An automated runner only works if you write clear card titles. If you create a card that simply says “solar?”, the agent will generate thirty pages of generic solar theory you never asked for. Good prompts produce good leverage; lazy prompts produce digital clutter.

Open Source Repository

The complete supervisor loop, worker contracts, and setup scripts are open source:

• GitHub Repository: github.com/jtothejz/personal-maestro
• License: MIT