What happened
As Product & AI Lead, my work comes in through mail, Slack, meetings and Jira, and so do other people's promises to me. Loops stayed open and follow-ups slipped. Too much of my head went to remembering instead of deciding.
I built an assistant that does the bookkeeping by itself and leaves me three verbs: send, approve, decide. It turns mail, meetings and quick notes from my phone into tasks, tracks what others owe me, checks whether a loop has really closed, drafts (never sends) follow-ups and plans focus time in my calendar. Everything runs on schedules; I don't trigger anything by hand. Once a week it prepares a review of my inbox and task list that I walk through in one sitting.
It runs as part of AI OS, with the same watchdog and guardrails.
How I built it
Google Tasks as the backbone. I built a small Python client, a CLI and a local MCP server (12 tools) so Claude can read and manage my to-dos. There is deliberately no delete: the assistant completes tasks, so anything can be reopened. Cloud routines can't use local MCP servers, so they install the CLI instead.
State lives in task lists, not in a database:
- "My tasks" for my own actions.
- "Waiting for" for what others owe me. Each item carries the date the question went out and, where the answer hangs on a meeting or a promised date, a "wait until" moment.
- An assistant log for deduplication and run markers.
- A capture list I fill from my phone.
Every task links back to its source (mail thread, Slack message, meeting) so the assistant can check it later.
Four cloud routines and two scheduled app tasks:
- Mail to task: a Gmail label becomes a task.
- Meeting to task: it reads meeting transcripts. My actions go to my list, other people's commitments go to "waiting for". Since 2 November it also processes my phone captures: each one-line note becomes a task, a decision log entry or a memory line, and the capture item records where it went.
- Waiting for: every weekday morning it asks one question per item: is this loop closed? It checks the evidence across mail, Slack, meetings and Jira. It closes the item with an evidence line, moves it to a real next step, or drafts at most two follow-ups. After that the item is marked stuck and I decide. Since 3 November it also tracks whether people confirmed the decisions I sent them, and shows silence as "not confirmed".
- Agenda planner: it ranks tasks by importance (not due date), checks first whether a task is already done, and plans at most two focus blocks a day. My own edits to the calendar always win.
- Monday review (a Claude app task, since 9 November): before my weekly usage reset it archives inbox mail that is provably handled, closes tasks that are provably done, and leaves the rest as a walk-through for me in that session. Nothing goes out.
- Watchdog: every weekday it checks that every scheduled job really ran, and messages me only when one didn't.
Two on-demand skills for the same job:
- Inbox cleanup (26 October): one agent per inbox thread checks whether the mail is handled (a sent reply, a ticket, a task, a note), archives what is done and reports each kept mail with a concrete next step.
- Task review (9 November): one read-only check per batch of tasks against mail, Slack, Jira, meetings and calendar, on a cheaper model first and a stronger model for every uncertain verdict. It closes what is proven done, puts obsolete items to me, then walks the rest with me in multiple-choice rounds: quick choices first, the harder ones collected and executed at the end.
Guardrails as code, not as prompts:
- Routines never send mail. Their only output is drafts plus a summary to my own Slack, and only when something changed.
- Write tools a routine doesn't need are blocked at the tool level.
- Everything the routines read is treated as data, never as instructions. Capture text never goes on a command line.
- Local hooks block mail without my signature, block sends that would carry the connector's mangled redirect links, and ask me to confirm calendar events before 09:45, over lunch or spanning days.
- A kill switch I can flip from my phone stops the meeting, waiting-for and agenda routines and the Monday review at their next start. The watchdog keeps running and tells me when the switch is off. The mail-to-task routine has its prompt built in, so it does not read the switch and is turned off on its own.
Draft-quality measurement, built and parked. I built a way to measure how much I change each draft the assistant prepares: a record per draft, an edit-size score and a replay over past sessions, with tests on synthetic data. It is switched off. When I stopped building improvements to the OS on 3 November, this one was frozen before it recorded anything.
Outcomes
- 77 open tasks cleared to zero on day one (22 October), all completed rather than deleted, then 15 real tasks added back. The list became trustworthy again.
- 484 tasks have passed through "My tasks" since 22 October: 444 completed, 40 open (13 November, including the 77 from day one).
- 88 items tracked in "Waiting for": 60 closed, 28 open (13 November).
- The first waiting-for run closed 4 loops and retitled 2. That was work I had been carrying in my head.
- The meeting routine's dry run over 10 meetings proposed 5 tasks with zero duplicates.
- The agenda planner's first run placed two focus blocks and annotated 25 tasks with when they would get attention.
- The watchdog's dry run produced 9 findings, all real.
- The core layer started on 22 October. The weekly review and the task-review skill followed on 9 November.
What I learned
"Did they reply?" is the wrong question. The first waiting-for routine only read mail, only counted replies from the person I was waiting on, and ignored what I had done since. Loops that were long closed stayed open. I rebuilt it around one question, is this loop closed?, with evidence from every channel.
Timing matters more than wording. It once drafted a follow-up on the morning of the very meeting where the answer was due. Items now carry a "wait until" moment.
An assistant can chase too much. I dropped three of four follow-ups in one review. The rule now: no chasing work a colleague owns or that sits in the running sprint. A heads-up mail gets no "waiting for" item at all.
Rank by importance, then check if it's done. The planner's first dry run picked by due date and scheduled a trivial task I had already finished.
Daily routines don't replace a weekly look. Even with four routines closing loops, 40 tasks and 28 waiting-for items stay open. Most need a choice only I can make. That is why the weekly review prepares the walk-through instead of trying to decide for me.
Measure only what you will look at. The draft-quality measurement was built and tested, then parked before its first record. Without a weekly moment to read the numbers it would have been one more log.
Connectors have quirks you only find in production. Mail search drops the dots in addresses. Links get wrapped in redirects. Updating a draft silently drops the signature. "Permitted tools" restricts nothing; only "disallowed tools" does. Each of these became a rule or a hook.
Not every experiment survives, and that's fine. I stopped a voice briefing over a phone line after a day (the platform couldn't call tools mid-call), and an in-app cockpit pane the same evening it was built. Both have write-ups of what went wrong.
How I did it
Solo with Claude Code, in small specs with a dry run before each one went live. Most routines' prompts live in git like code, every change to a schedule is recorded, and a watchdog checks on the watchers. The principle throughout: features run on their own, and the human only sends, approves and decides.