Cosa contiene

Un file SKILL.md (la guida operativa) + un modello di partenza. Un'anteprima:

Background Task Budget Planner

Why this exists

A proactive agent's background work counts against the same usage limits as your own chats. Duet says so plainly: background checks and larger tasks draw on your plan allowance, and you can reach your limits "sooner than with chat alone." Nobody budgets for that — until the week the agent's morning briefs cost more than the work they saved.

Workflow

1. Measure a normal week (input: plan limits + last week's agent activity -> output: a baseline).
   Record: number of check-ins per day, average tasks per check-in, and one
   large task (a dashboard, a forecast, a web app) with its cost if you can
   see it. If the product does not show per-task cost, measure by proxy:
   how much of the week your limit meter moved on days the agent ran vs
   days it did not.
2. Tier the work (input: your task list -> output: 3 tiers).
   Tier 1 — cheap and scheduled: briefs, inbox/calendar scans, invoice
   watches. Tier 2 — medium and bounded: drafting, research summaries,
   dashboards. Tier 3 — expensive and rare: building apps, financial
   models, multi-app workflows. Each tier gets a weekly cap (e.g. Tier 3:
   one task per week, only on a named day).
3. Set the cadence (input: urgency ladder -> output: check-in schedule).
   Not everything needs checking every few hours. Split sources into:
   every check-in (calendar, priority inbox), twice daily (general inbox,
   payments), daily (social, news). An urgent message can sit until the
   next check-in — so the cadence for a source must match how long you can
   actually tolerate not knowing.
4. Write kill criteria (input: tier caps -> output: stop rules).
   A task stops when: it exceeds its tier's time/cost estimate by 2x, it
   retries the same failing step three times, or it asks for a new
   permission mid-task. Stopped tasks report what they did and what they
   did not finish — a silent stop is a failed task, not a saved budget.
5. Weekly review (input: the week's meter -> output: one adjustment).
   Each week, change exactly one thing: a cadence, a cap, or a tier
   assignment. If the meter moved less than expected, the agent is not
   checking enough to be useful; if it moved more, something is Tier 3 in
   Tier 1 clothing.

Rules of thumb

• A brief that costs more than the 5 minutes it saves is a hobby, not a tool.
• Background work should never exceed one third of your weekly allowance;
  reserve the rest for work you initiate.
• Never let the agent's own usage dashboard be the only record — copy the
  weekly numbers into your own note. Product dashboards change; your note
  does not.

Anteprima del modello

Background Task Budget Worksheet

1. Baseline week

• Plan limit:
• Check-ins per day:
• Large task + cost (if visible):

2. Tiers

3. Cadence

4. Kill criteria

• Cost/time overrun:
• Retry limit:
• New permission mid-task:

5. Weekly adjustment (one change only)

-

Il pacchetto della skill (SKILL.md + modello) è scritto in inglese.

Acquista il pacchetto completo della skill — $4.99 Pro — tutte le 511 skill