← Library · AI-OS cookbook · 07
A two-surface system: a weekly focus plan that frames the week, and one fresh task list per week — capped at five items — that freezes into a paper trail. Plus an AI triage agent that routes every to-do dump against your anchors and shows its reasoning, so loops close instead of carrying over forever.
One ever-growing to-do list. Everything carries over by default, strategic work and ten-minute errands fight in the same pile, the plan is just todos — and nothing records what was actually prioritized when.
A weekly focus plan frames the week; a fresh "Week NN" list holds at most five items that ship this week; old lists freeze as the historical record; long-arc work lives apart — and an AI agent triages every dump, routes it to the right surface, and reports its reasoning for override.
Planning collapses when strategy and execution share one list. Keep them apart: a focus plan note that carries the week's anchors and decisions, and a weekly task list that carries only what ships in the next five days.
A weekly note ("Week NN — focus plan") in your notes app: the week's anchors, the principles, the current state. It frames every triage decision — it is not a to-do list, and tasks never replace it.
One Google Tasks list per ISO week, named Week NN. It contains only items that ship this week. The week is the unit of execution: one week to get things done.
When a new week starts, the old list stays exactly as it is. Never delete it: the frozen lists are the historical record of what was prioritized each week — self-dating, no renaming needed.
The default is don't carry. Whatever didn't ship drives away, and the new week gets re-thought against what matters now. An item earns its place in Week N+1 on the new week's shape — not on being leftover.
Anything longer than a week — 30-day milestones, backlog, signature long-arc work — belongs in the default list, not the weekly one. That protects it from the weekly reshuffle and keeps the weekly list honest.
Max 5 active items. More and the list stops working at a glance. Sort ★ → A → B, the one thing on top. Short titles, detail in a one-line note. Delegated items you only monitor go in a separate tracks list — the weekly list is your own execution only.
"Each week, all the cards get reshuffled. The next week can carry parts over, but new things land too. What we don't do is drag everything across every single week. I have one week to get things done — whatever I didn't finish drives away first, and we re-think what matters this week. And I want the history as a paper trail: a new list, every week."
The operator's job is to surface what feels important — talking, mid-meeting, mid-walk. The agent's job is to assess each item with everything it knows and decide where it lands. Transparently, so every call can be overridden.
Items arrive as they come — spoken, typed, half-formed. No prioritizing at capture time; capture is cheap, sorting is the agent's job.
Each item is held against the current focus plan's anchors (your 3–4 strategic lines plus family/personal). Aligned with none of them → low priority: backlog or drop.
If this week's list is already full of A-items, a new item defaults to B or punts to next week — unless it's strictly time-sensitive. If the item depends on something not yet done, the agent flags the precondition explicitly so the dependency is visible.
Every item lands in exactly one place:
| Surface | When it lands there |
|---|---|
| "Week NN" list | Ships this week and fits the cap |
| Default (long-arc) list | More than a week out — milestones, backlog |
| Calendar block | Only if the work needs ≥60 minutes of protected time and the slot is conflict-free. Default is task-only — the operator slots it to the calendar if wanted |
| Tracks list | Delegated — someone else drives it, the operator only monitors |
| Drop | Not strategically aligned — and the agent says so out loud |
Every triage returns: what was added where (list, priority, due date), why (which anchor, why this priority, what precondition), and what was deliberately not done. Ask "what did you drop?" and the agent replays every item with its one-line reasoning — so individual calls get overridden without redoing the pile.
"Treat what I tell you as a to-do dump. For each item: assess strategic alignment against my current focus-plan anchors, check this week's capacity and any preconditions, then route it — Week list, long-arc list, calendar block, or drop. Be transparent: tell me what you added where and why, what you deliberately didn't do, and where I should challenge your read. When I ask what you dropped, give me every item with its one-line reasoning so I can override individual calls."
A plan that's just todos changes nothing. The focus plan exists to compensate for three structural leaks most operators share — so its sections are mandatory, not nice-to-have.
An explicit list: who · when scheduled · what to decide · briefing notes ready before the meeting. This forces expectations onto the table and puts the hard conversations on the calendar — with prep, so they actually happen.
Every open question structured as: 3 options (hard cap) · a recommendation · an owner · a decision-by date. Never collect thirty open questions — cap the options and force the date. This is the antidote to option-collecting.
Front-load what only the operator can do this week — stages, key commercial conversations, signature content. Everything else becomes a delegation with owner · ship date · definition of done. The acceptance criteria are the point: "great" stops living in one person's head.
Before writing the plan, the agent pulls what actually happened: recent meeting transcripts, team-channel threads, last week's planning notes. Then it surfaces the delta since the last plan — what changed, what slipped, what's newly urgent.
Anything untouched for 3+ weeks — a client line, a project, a squad — goes on the decision table explicitly: revive it, delegate it, or kill it. Dormant ≠ decided.
First: true up the old list — mark anything that's actually done, so the paper trail is accurate. Then freeze it and create the new Week NN list. The agent pulls recent meetings, team threads and planning notes to spot new commitments, then proposes the week: new commitments + deliberate carries + fresh personal resets.
Don't pad. Fewer items, sharper — five is the ceiling, not the target.
The week's top-5 get re-read every morning (they live where you'll see them — see the Calendar OS, cookbook 06). Through the week, to-dos get dumped as they surface; the triage agent routes them in real time and the list stays at five.
Whatever didn't ship drives away by default. The next week is re-thought from zero against the anchors; an unfinished item must re-earn its place. The frozen lists pile up into the one thing an endless to-do list never gives you: a history of what you chose.
Copy the bootstrap promptThe button below puts it on your clipboard.
Paste it into Claude CodeIn a session with Google Tasks and your calendar connected via MCP.
Answer its questionsIt sets up the surfaces, installs the triage protocol, and runs your first weekly cycle.
No download needed — this cookbook is the system + an AI triage agent you run on your own stack.