How I work with AI

An assistant that runs in the background is only as trustworthy as the rules it runs under. Here are the rules, then the system that enforces them, then the tools.

Principles

  1. It never sends anything to another person. Drafts stay drafts until I send them.
  2. It never closes my decisions for me. It recommends; I mark done, decline, approve.
  3. It never spends money. It can comparison-shop and pressure-test a purchase. The card stays with me.
  4. It writes down what it learned. Every project has a book: brief, decisions, open questions, session log. The delta between what it drafts and what I send is how it learns my voice.
  5. It brings me decisions, not updates. Recommendation first, evidence underneath, prep already done.

The system

Me (decides, sends, buys) Scheduled Claude run Notion, the system of record Outside service

The loop

Everything runs on one loop. Scheduled runs read the outside world and the Notion record, do the work an agent is allowed to do, and leave decisions in an attention queue. I read the queue from one page, answer in a line, and that answer starts the next run. Nothing leaves the system to another person unless I send it myself.

Tyler decides · sends · buys Command Center page one screen, fixed places, 5-second read Notion (system of record) Attention Queue Needs Tyler · Agent working · Blocked Project Books (one per project) Radar tables (jobs, concerts, shipping) Scheduled Claude runs 7:53am · 3:53pm · daily radars plus any run a reply starts drafts, researches, files, recommends Gmail Calendar Drive Spotify GitHub reads, replies in a line queue every 15 min mark done, comment "Send to Claude" starts a run now writes queue items, books, tables reads the book first read only Human gate drafts never sent · nothing bought decisions never closed for me
One loop: runs read outside services and the Notion record, write back into Notion, and surface decisions on one page. Replies from that page start the next run.

Project Books: how it remembers

Every project, from a renovation to a job search to a dinner planner, gets one Notion page with the same four parts. A run reads the book before it does anything and appends a session log line when it is done. That is the whole memory model: no hidden state, and I can read the book instead of the transcript.

Project Book · one page per project 1. Briefgoal, constraints, who 2. Decision logdated, with the why 3. Open questionsnumbered, owner named 4. Session logone line per run Subpages: prep docs, story bank, call notes, drafts, data tables (linked, not pasted) Scheduled run start: read 1-4 and the queue end: append to 2 and 4 reads first appends last Tyler approves changes to the brief answers open questions gated edits Command Center queue items link to their book
Memory is a page, not a transcript: a run reads the book before it acts and appends to it when it finishes. Changes to the brief and answers to open questions are mine.

The learning loop: the delta between the draft and what I sent

The agent never sends mail. It writes a draft in my voice and leaves it in Gmail; I edit and send. On the next run it compares what it drafted to what actually went out, and folds the difference into a writing-style file that every future draft reads. The same shape runs everywhere: Job Radar notes revise the rubric, project books carry decisions forward, and the Command Center queue records which recommendations I took.

Writing-style profile openings, tone markers, sign-offs, words to avoid, who to cc Run drafts the email reads the profile, writes a Gmail draft, nothing more reads Gmail draft, unsent saves Tyler edits, then sends from his own account Next run compares sent message vs. the draft what did he change, and why reads sent mail proposes an update; Tyler approves Same loop elsewhere: Job Radar calls → rubric · queue answers → project book decision log · purchase questions → purchase advisor
The correction loop: draft, human edit, diff, profile update. The assistant gets better at sounding like me without ever sending as me.

The per-project diagrams, for the Job Radar and Concert Radar, are on their project pages.

Tools

Click anywhere or press Esc to close.