Skip to content

983-zao-assistant-todo-workflow

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

983-zao-assistant-todo-workflow

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

topic dev-workflows
type guide
status research-complete
last-validated 2026-07-06
superseded-by
related-docs 601, 888, 973, 978
original-query Research the ZAO assistant + todo workflow we built this session and what more we can do with it - one Supabase tracker as source of truth -> 3 doors (Claude/ZOE/board), auto-tagging, delegation lanes, research-grind loop, clipboards, top-3 surfacing. Research external best practices and produce a prioritized improvement backlog.
tier STANDARD

983 - The ZAO assistant + todo workflow: what we built, what to build next

Goal: We built a real personal-assistant workflow this session (one tracker, three doors, auto-tagging, delegation lanes, a research-grind loop). This doc grounds it against 2026 best practice and hands back a prioritized backlog of what to build next - so the system keeps up with how Zaal actually works (many parallel agent threads, scarce human judgment).

Recommendations (build in this order)

# Build Why Effort Value
1 Owner-axis on tasks (me / agent / review / blocked) The #1 finding: with parallel agents the bottleneck is judgment, not execution. The board shows status but not "what needs YOU vs what an agent can take." MED HIGH
2 Weekly review routine (ZOE cron, Sunday) GTD's "reflect" step catches the silent failure - a task with no next action or no date. We have 52 undated tasks right now; nothing surfaces the stale ones. LOW HIGH
3 ZOE morning briefing pulls top-3 + "decisions waiting on you" ZOE already has a Morning Briefing routine; wire the tracker's top-3 + review-queue into it so the surfacing is automatic, not hand-run. LOW HIGH
4 Auto-tag on creation (classifier at the write path) We batch-tagged 54 items once; new tasks land untagged. Classify brand/theme/work-type when any door creates a task. MED MED
5 Auto-route research-queue -> research loop 4 items are tagged delegated_to=research-queue but nothing picks them up. A scheduled grind should drain that lane into docs (like 974-980). MED MED
6 Theme filter chip in the board UI The auto-tagger writes metadata.themes (web3/ai/music...) but the board can't filter on it yet. Small UI change closes the loop. LOW MED
7 HITL checkpoint classes on actions (trivial / reversible / irreversible) Formalize what the assistant already does ad-hoc: soft-archive = reversible (auto-OK), PR-merge-to-live = irreversible (gate on Zaal). MED MED

What we built (grounded in the system)

One source of truth -> three doors. Every task lives in the Supabase cowork tracker (tasks table, project etwvzrmlxeobinrlytza). Three interaction surfaces read/write it:

  1. Claude terminal (this session) - creates, tags, archives, drafts via REST.
  2. ZOE Telegram (@zaoclaw_bot) - notifications + capture; has live routines (Morning Briefing, Lunch Ping, Nightly Processor).
  3. The board app (ZAODEVZ/ZAOcowork, thezao.xyz) - getActions() in src/lib/data.ts loads the board; TaskRoom.tsx is the rich detail (status, priority, due, comments, dependencies); Board.tsx has filters + saved views. As of this session (PR #118) personal items fold into /board + /my-work.

Auto-tagging. A keyword classifier tagged all 54 active items with brand (brands[], filterable now), work-type (category, the board's canonical enum), and theme (metadata.themes: web3/ai/music/events/growth/governance/research).

Delegation lanes. Items route off the active list via metadata.delegated_to: a dedicated fractal build terminal (4 items), the ZOL agent (1), a research queue (4). Plus a bucket=later lane (4 items) and the archive.

The research-grind loop. A self-paced loop shipped 7 docs (974-980) in this session, one per tick, each: reserve number -> write -> secret-scan -> PR -> tracker update -> ZOE ping. Human-out-of-the-loop for the research, human-in-the-loop for the merge.

What 2026 best practice says (external)

The strongest, most consistent finding across six sources: the constraint has moved from execution to judgment. GTD assumed one human, one task, serial attention; with parallel agents "the next action is deciding which thread gets your judgment next" (Jay Schulman). The to-do list splits into three lanes GTD never had: threads in flight, decisions waiting on you, outputs to review. The review queue is "the new inbox, and it fills faster than the old one ever did."

Concrete patterns worth adopting:

  • Next-action-owner axis (Personal Task Assistant, dev.to): tasks carry both status AND an owner (me / agent / unassigned), and the value is the filtered queue - "what do I do / what can the agent do / what's ready for review / what's blocked / what's overdue." Their agent-queue endpoint returns a summary: {active, overdue, due_soon, agent_ready, human_input, review, blocked}.
  • HITL checkpoint classes (Knowlee): the agent classifies its own actions as trivial (auto-approve, log only), reversible (approve in batches, async, even after the fact), or irreversible (approve synchronously, before acting). "Review the irreversible class carefully, the reversible class quickly, the trivial class never."
  • LLM-as-GTD-maintainer with review cadences (stakiran gist): the human captures/decides/executes; the LLM owns the maintenance - sorting, retagging, and running daily / weekly / monthly / yearly reviews. The weekly review's job is to catch "the project with no next action - the silent failure that rots the whole system."
  • One bounded step per agent + run record (cloudbuddyapps): each agent owns one repeatable move (classify/extract/draft/route), sensitive actions stay gated, and every run logs what it saw/did/who reviewed.
  • Classifier router (simaba/agent-orchestration): a classifier agent routes tasks to the right specialist by declared capability - the pattern behind our auto-tagger + delegation lanes.

Gap analysis (our system vs the patterns)

Pattern Do we have it? Gap
Single source of truth YES - the tracker none
Multi-door capture YES - Claude/ZOE/board none
Auto-classification PARTIAL - batch only not on the write path (Rec #4)
Delegation lanes YES - delegated_to no auto-pickup for the research lane (Rec #5)
Owner-axis (me/agent/review) NO - only status + owner_id the biggest gap (Rec #1)
Review cadence NO 52 undated tasks, nothing surfaces stale (Rec #2)
Top-3 / decisions-waiting surfacing PARTIAL - hand-run not wired into ZOE's routine (Rec #3)
HITL checkpoint classes INFORMAL not encoded (Rec #7)
Run record / audit trail PARTIAL - task notes good enough for now

Also See

  • Doc 601 - the 5-surface cleanup this workflow sits on top of.
  • Doc 888 - ZOE reliability/routing (the door that would run Rec #2 and #3).
  • Doc 978 - the numbers-framing doc (velocity metrics would extend it).

Next Actions

Action Owner Type By When
Add a next_owner field (me/agent/review/blocked) to the tracker + surface it as a board lane and a /my-work split @Zaal PR 2026-07-20
Ship a ZOE weekly-review cron (Sunday AM): list undated P1s, tasks with no next action, stalled >14d, and ask Zaal to date/re-scope each @Zaal Bot task 2026-07-20
Wire the tracker top-3 + "decisions waiting on you" into ZOE's existing Morning Briefing routine @Zaal Bot task 2026-07-13
Add a classifier call to the task write-path so new tasks auto-tag brand/theme/work-type on creation @Zaal PR 2026-07-27
Add a board filter chip for metadata.themes so web3/ai/music are filterable in the UI @Zaal PR 2026-07-27
Schedule a research-grind loop to drain delegated_to=research-queue into ZAOOS docs (like 974-980) @Zaal Loop 2026-07-13

Sources