---
name: weekly_review
description: Run a GTD-lite weekly review for AI agents in 30 minutes: groom backlogs, audit open commitments and waiting-fors, and plan the week ahead. Triggers on weekly review requests, cadence-based review schedules, or when an agent's task lists and promises have drifted out of sync.
---

# Weekly Review

## Purpose
Keep the agent's entire commitment system trustworthy. Once a week, spend 30
timeboxed minutes getting clear (process all inputs to zero), getting current
(every list reflects reality, every open loop has a next action), and getting
creative (parked ideas get a fresh look, new ideas get captured). Adapted from
David Allen's GTD Weekly Review — the canonical 11 steps, translated from
human desk-work to agent work queues. Method source:
`references/phases.md`.

## Workflow
1. **Open a review report.** Run
   `python3 scripts/weekly_review.py new --title "<YYYY-MM-DD> weekly review"`
   to emit the skeleton, or copy `references/review-template.md` by hand. Work
   inside the report for the whole session.
2. **GET CLEAR (8 min).** Sweep every capture surface into the report's inbox
   section: unread threads and messages, pending tool/browser-task results,
   downloads and artifact folders, approval requests, stray notes. Then clarify
   each item until the inbox is empty: delete/no-op, do now (<2 min), next
   action, project, waiting-for, someday/maybe, or file as reference. Finish
   with a head-sweep: write down every open loop you can think of and clarify
   those too.
3. **GET CURRENT (15 min).** Walk the six lists in order:
   next-actions/tracked items → previous 7 days (promises kept? slippage?) →
   upcoming 7 days (schedules, deadlines, renewals, prep work) →
   waiting-for (nudge what's nudgeable, re-date the rest) →
   projects/goals (each gets a status and ≥1 next action; park or kill the
   dead) → standing checklists/runbooks (still current?).
4. **Commitment audit.** Every promise made to the principal in the last 7 days
   gets one row in the audit table: commitment, made-to, made-when, due,
   status, next action. Anything without a next action fails the review.
5. **GET CREATIVE (7 min).** Review someday/maybe: activate what's viable,
   delete what's dead. Then capture new ideas, risks, experiments — uncensored.
6. **Close out.** Mark each of the 11 steps done in the report, write the
   one-line verdict (clear / current / creative: pass or fail with a note),
   and run `python3 scripts/weekly_review.py check <report.md>` — it fails if
   any step is unmarked or any audit table has an empty next-action cell.
7. **File it.** Save the report where future-you can find it (dated log or
   goal workspace). Carry unfinished next-actions back into the live system —
   a review that doesn't update the real lists didn't happen.

When the budget is spent, move on: a partial honest review beats a skipped
perfect one. Run every 7 days; give it at least 4 consecutive runs before
judging the habit.

## Output Contract
One dated markdown report containing, at minimum:
- Inbox-zero confirmation (what was captured, how each item was disposed)
- Commitment audit table: commitment | made to | made when | due | status | next action
- Waiting-for table: item | blocked on | since | nudge sent? | next follow-up
- Project health table: project | status | next action | park/kill/continue
- Next-7-days plan: date | item | prep needed
- Someday/maybe deltas (activated / deleted)
- Creative captures (raw, unjudged)
- Verdict line: `clear: pass|fail — current: pass|fail — creative: pass|fail — <note>`
- All 11 GTD steps marked done (the `check` script enforces this)

## Operating Rules
1. **Timebox strictly.** 8 / 15 / 7 minutes. When a phase's budget is spent,
   record where you stopped and move on.
2. **Verify, don't invent.** Never mark something done unless a tool result,
   file, or message confirms it. Unverifiable = marked `unverified`, never
   `done`.
3. **Nothing silently disappears.** Deleted or killed items get one line: what
   it was, why it's gone.
4. **The review serves the system, not an audience.** Be honest about slippage;
   a clean-looking report built on fiction is worse than an ugly honest one.
5. **Update the real lists.** The report is a snapshot; the task lists, tracked
   items, and schedules are the system. If the review didn't change them, redo
   step 7.