The Weekly Scheduled-Task Audit: Make Your AI Workforce More Effective Every Week

By Dennis Yu

A weekly scheduled task audit is the management layer for an agent workforce. When a company starts using agents seriously, the work moves from one-off requests into ongoing functions. Marketing gets a content queue. Sales gets follow-up and pipeline reviews. Operations gets reporting, quality assurance, and client delivery checks.

That is what a healthy organization should do. A business does not grow by remembering to ask for the same work every Tuesday.

But scheduled jobs create a second problem: task sprawl. The list gets longer. Prompts overlap. Three jobs pull the same data. One job starts before its source is ready. Reports get produced because the schedule says so, not because anyone makes a decision from them. Token usage rises while nobody can say which jobs generated traffic, leads, revenue, retention, or reusable knowledge.

So the agent workforce needs a manager, too.

This article gives you a weekly portfolio audit you can run in Claude Cowork, ChatGPT Work, Codex, or any capable agent system. The audit reviews all your recurring jobs as one operating system, then tells you what to invest in, combine, sequence, instrument, reduce, pause, or retire.

The governing rule: effectiveness comes before efficiency. Saving 40% of the tokens on a job that does not move the business is not a win. Spending more on a job that reliably creates qualified calls may be exactly right.

What the weekly audit is supposed to accomplish

The audit answers five questions:

1. What exists? Build one inventory of scheduled jobs across marketing, sales, operations, finance, and client delivery.
2. What depends on what? Map shared inputs, duplicate pulls, upstream data refreshes, downstream jobs, owners, and handoffs.
3. What creates value? Connect each job to traffic, qualified leads, booked calls, revenue, retention, risk reduction, or repeatable leverage.
4. What should change? Recommend a clear disposition: invest, keep, consolidate, sequence, reduce, instrument, pause, or retire.
5. What is missing? Produce the shortest possible dependency list—connectors, exports, permissions, tracking, owners, or baselines—needed for a more truthful review next week.

This is MAA: Metrics, Analysis, Action applied to your entire agent workforce. Metrics say which jobs ran and what they cost. Analysis explains why the business result moved. Action changes the portfolio.

Why you should audit the portfolio, not each prompt in isolation

Most optimization happens one prompt at a time. That misses the largest savings and the biggest failure modes.

Imagine these five jobs:

• A Monday SEO report pulls Search Console and GA4.
• A Tuesday content-prioritization job pulls the same two sources.
• A Wednesday client report pulls them again.
• A daily publishing agent creates posts before the content inventory finishes refreshing.
• A Friday leadership brief summarizes all four jobs without checking CRM revenue.

Each prompt can look reasonable on its own. The portfolio is still badly designed.

A portfolio audit can replace repeated data pulls with one verified weekly snapshot, run the inventory refresh first, let downstream jobs read the same artifact, and add the missing CRM outcome data to the leadership brief. That is more than token efficiency. It makes the conclusions more consistent and the work more useful.

The same logic applies to the Content Factory. Produce, Process, Post, and Promote are not independent chores. They are a system. Plumbing comes before production; Perform feeds results back into the next cycle. Your scheduled jobs should follow the same dependency graph.

The minimum task registry

The first run should create a task registry. A spreadsheet, CSV, database table, or Markdown file is fine. Use one row per recurring job and capture these fields:

If a field is unknown, write unknown. Do not replace missing data with a confident estimate. A labeled estimate can be useful; an unlabeled guess poisons the whole audit.

The task registry becomes business truth. Runtime logs tell you what executed. The registry tells you why it exists, what it costs, who owns it, what it influences, and how it should be evaluated. This is the same reason we maintain an Asset Tracker and teach agents to improve their own instructions through meta articles.

Score value without pretending attribution is perfect

Do not collapse everything into a single magic number on the first run. Keep effectiveness, efficiency, and confidence visible as separate dimensions.

1. Effectiveness score

Normalize each component to 0–100 and use this starting rubric:

Change the weights when your business model demands it. A local service business may weight booked calls more heavily. A compliance operation may weight risk reduction. A new content property may temporarily weight qualified traffic and indexation.

2. Efficiency score

Track efficiency beside impact:

• Cost per useful output
• Tokens or credits per run
• Tool/API units per run
• Human review minutes per run
• Time from source-ready to decision-ready
• Duplicate data pulls avoided
• Rework and retry rate

Efficiency helps you choose how to run valuable work. It should not decide whether the work matters.

3. Evidence confidence

Give every conclusion a confidence grade:

• A — Direct: CRM revenue, booked-call record, experiment result, or verified system log.
• B — Strong proxy: Qualified lead, assisted conversion, accepted deliverable, or verified downstream use.
• C — Weak proxy: Traffic, engagement, completion count, or human time saved without an outcome link.
• D — Assumption: Anecdote or estimate with no connected source.
• U — Unknown: There is not enough information to judge.

A task with high apparent impact and D-grade evidence should usually be instrumented, not automatically expanded. A task with low vanity metrics but direct evidence of retention or revenue may deserve more effort.

Use the Evidence Verification skill when the agent must separate verified facts from claims and estimates.

The eight decisions the audit can make

Every job gets one primary disposition:

The audit should also identify missing functions. If nobody checks whether Content Factory articles get indexed, whether calls are qualified, or whether client reports lead to action, the answer is not always to cut a job. Sometimes the highest-value recommendation is a new measurement task.

What the agent may change on its own

Autonomy needs an explicit boundary.

The auditor may automatically:

• Update the registry, dependency map, and baselines.
• Reuse an existing verified snapshot instead of pulling the same source again.
• Reorder work inside the audit itself.
• Improve report formatting and remove duplicated analysis.
• Create a draft consolidation plan, revised prompts, and proposed schedules.
• Make a reversible, internal, low-risk change when the policy explicitly allows it and a rollback is recorded.

The auditor should request approval before it:

• Deletes, pauses, or retires a scheduled job.
• Changes a client-facing cadence or output.
• Merges jobs with different owners, permissions, or compliance requirements.
• Sends email, publishes content, changes CRM records, edits campaigns, or contacts a person.
• Changes a budget, bid, production system, or source of truth.
• Makes a structural change that could break downstream work.

This is consistent with how we run the Google Ads MAA Agent: the agent pulls, diagnoses, and stages the work; a person reviews the consequential action. The Weekly Brand MAA canonical SOP uses the same pattern—verified sources, safe additive changes when permitted, and human review for structural work.

The five-page weekly decision pack

Ask the audit to produce a five-page PDF, not a 50-page activity dump.

Page 1 — Executive decision view

• Total active jobs, runs, cost, failures, and human review time
• Top five value creators
• Bottom five or unknown-value jobs
• Decisions requiring approval
• Biggest missing dependency

Page 2 — Organization map

• Marketing, sales, operations, finance, and client delivery lanes
• Human owners and virtual workers
• Inputs, outputs, and handoffs
• Orphaned jobs with no owner or consumer

Page 3 — Dependency and consolidation map

• Upstream data refreshes
• Downstream consumers
• Duplicate pulls or duplicate analyses
• Jobs that should be combined
• Jobs that should run in sequence

Page 4 — Value, cost, and confidence

• Effectiveness versus cost matrix
• Revenue, lead, traffic, impact, leverage, and reliability view
• Evidence confidence for every major conclusion
• Trend against last week and baseline

Page 5 — Action plan and dependency requests

• Changes completed safely
• Changes staged for approval
• Owner and due date for each action
• Exact connectors, permissions, exports, tracking fixes, or decisions needed
• What the next run will be able to answer after those dependencies are supplied

The PDF is a decision surface. The registry remains the detailed source of truth.

Copy-paste master prompt

Prefer a reusable file? Download the portable Weekly Scheduled-Task Portfolio Audit skill, unzip it, and add SKILL.md to your agent project or skill pack. The complete prompt remains below so anyone can use the method without installing anything.

Paste the prompt below into a new agent task. Replace the bracketed fields, test it once manually, then schedule it weekly after your analytics and CRM data have refreshed and before your leadership review.

You are the weekly portfolio manager for our scheduled AI jobs and recurring automations.

GOAL
Maximize business effectiveness first, then efficiency. Review the entire portfolio as an operating system—not as isolated prompts. Determine which jobs create traffic, qualified leads, booked calls, revenue, retention, risk reduction, or repeatable leverage; which jobs should receive more effort; which should be consolidated, sequenced, reduced, instrumented, paused, or retired; and which important functions are missing.

PARAMETERS
- Organization: [NAME]
- Review window: [LAST 7 DAYS]
- Timezone: [TIMEZONE]
- Departments: [MARKETING, SALES, OPERATIONS, FINANCE, CLIENT DELIVERY]
- Scheduled-task sources: [CLAUDE / CHATGPT / CODEX / CRON / ZAPIER / MAKE / PROJECT MANAGEMENT SYSTEM]
- Knowledge sources: [PROJECT FILES / DRIVE / OBSIDIAN / NOTION / WEBSITE / META ARTICLES / SOP LIBRARY]
- Outcome sources: [GA4 / GSC / CRM / CALL TRACKING / AD PLATFORMS / ACCOUNTING / SUPPORT / PROJECT MANAGEMENT]
- Cost sources: [MODEL USAGE / TOKEN LOGS / API UNITS / CLOUD COST / HUMAN REVIEW TIME]
- Registry location: [PATH OR URL]
- Prior reports and baseline: [PATH OR URL]
- Delivery location: [PATH, DRAFT DESTINATION, OR REVIEW THREAD]
- Autonomous-change policy: [STAGE ONLY by default]

NON-NEGOTIABLES
1. Effectiveness is more important than efficiency.
2. Business outcomes beat activity and vanity metrics.
3. Never fabricate missing metrics. Label estimates, assumptions, and unknowns.
4. Separate effectiveness, efficiency, and evidence confidence.
5. Follow Metrics -> Analysis -> Action. Explain why before recommending what to do.
6. Reuse verified source snapshots when several jobs need the same data.
7. Preserve permissions, owners, audit trails, and rollback paths.
8. Do not delete, pause, retire, merge, publish, send, edit campaigns, change budgets, or modify external systems without explicit approval unless the parameter above grants that exact action.
9. Prefer reversible internal changes. Record every automatic adjustment and its rollback.
10. If a source is unavailable, continue the audit, lower confidence, and put the smallest exact unblock request on the dependency list.

RUNBOOK
A. INVENTORY
- Enumerate every active, paused, and recently failed scheduled job you can access.
- Capture: ID, name, department, owner, schedule, timezone, purpose, inputs, outputs, upstream dependencies, downstream consumers, model/tools, average tokens or credits, runtime, human review time, success rate, last run, next run, and delivery channel.
- Compare with the existing registry. Add missing jobs and flag registry entries that no longer exist.

B. OBSERVE USAGE
- For each job, inspect recent run logs and whether its output was opened, accepted, used downstream, published, sent, assigned, or acted upon.
- Identify retries, duplicate sends, stale inputs, silent failures, unused outputs, and repeated manual corrections.
- When command-line or telemetry data exists, measure actual throughput, duration, cost, and error rate.

C. CONNECT TO OUTCOMES
- Trace each job to the strongest available business outcome: revenue or retention first; then qualified leads or pipeline; then attributable traffic or discovery; then strategic impact, risk avoided, or time saved.
- Record the source, attribution method, measurement window, and confidence grade: A direct, B strong proxy, C weak proxy, D assumption, U unknown.
- Do not reward a job merely because it ran or produced a deliverable.

D. MAP THE SYSTEM
- Build a dependency graph across marketing, sales, operations, finance, and client delivery.
- Identify duplicate data collection, duplicate reasoning, incompatible sources of truth, race conditions, schedule collisions, jobs that trigger before inputs are ready, and outputs with no consumer.
- Find opportunities for one upstream verified snapshot to serve multiple downstream jobs.
- Find opportunities to sequence jobs rather than have one agent poll or trigger another unnecessarily.

E. EVALUATE
- Score effectiveness using this starting rubric: revenue 30%, qualified leads/pipeline 20%, traffic/discovery 15%, strategic or operational impact 15%, repeatable leverage 10%, quality/reliability 10%. Adjust only when the business model clearly requires it and explain the change.
- Evaluate efficiency separately: cost per useful output, tokens/credits, tool units, human minutes, latency, duplicate pulls, retries, and rework.
- Compare the last 7 days with the prior 7 days and the original baseline when available.
- Be impartial, candid, and maximally truthful. Show uncertainty and conflicting evidence.

F. DECIDE
- Assign one primary disposition to every job: INVEST, KEEP, CONSOLIDATE, SEQUENCE, REDUCE, INSTRUMENT, PAUSE, or RETIRE.
- For consolidation, name the shared upstream step, what remains distinct, expected savings, risks, owner, and rollback.
- For sequencing, name the producer, artifact, freshness requirement, consumer, and proposed timing.
- For model choice, use the least expensive model that has demonstrated acceptable quality for that step. Use stronger models for judgment-heavy diagnosis when their incremental quality matters. Do not lower quality merely to reduce tokens.
- Suggest missing high-value jobs where the portfolio has a measurement, quality, handoff, or follow-up gap.

G. ACT SAFELY
- Automatically update the registry, baseline, dependency map, and internal audit artifacts.
- You may make only reversible, internal, low-risk adjustments explicitly allowed by the autonomous-change policy.
- Stage all consequential changes for approval with an exact before/after diff.
- Never treat a recommendation as implemented unless you verify the changed system.

H. DELIVER
Create:
1. An updated machine-readable task registry.
2. A full Metrics -> Analysis -> Action report.
3. A one-page approval queue with exact before/after changes.
4. A dependency checklist that is easy to complete: system, access or artifact needed, why it matters, owner, exact steps, and what becomes measurable afterward.
5. A five-page visual PDF:
   - Page 1: executive decision view
   - Page 2: organization map with human and virtual workers
   - Page 3: dependency, consolidation, and sequencing map
   - Page 4: value versus cost with evidence confidence
   - Page 5: action plan, owners, due dates, and dependency requests

FIRST-RUN RULE
The first runs will have missing access and weak baselines. Do not stall. Produce the best honest view possible, mark unknowns, and rank dependency requests by expected decision value divided by setup effort. Make each request executable in under 10 minutes when possible.

SELF-IMPROVEMENT
At the end of every run, append: what was missing, what caused rework, which estimate can become a direct measurement, and one proposed improvement to this task or its skill. Do not silently rewrite the method; stage material instruction changes with a diff.

Turn it into a scheduled job

Claude Cowork

Create a Cowork project for the audit, attach the registry and this skill, connect the systems it may read, then use /schedule or the Scheduled page. Claude’s current help documentation says scheduled Cowork tasks can use connected tools, skills, and installed plugins. Start with stage-only approval mode and a weekly cadence. See Anthropic’s official guide to scheduling recurring tasks in Claude Cowork.

ChatGPT Work or Codex

Test the prompt in a normal task first. Then create a weekly scheduled task in the same project or conversation so the audit can reuse the registry and prior decisions. OpenAI’s current guidance recommends testing the first runs and keeping durable instructions in the prompt or an attached skill; local-project runs need the computer and desktop app available. See the official guides to scheduled tasks, building skills, and auditing a workflow.

Any other agent or automation system

Store the prompt and SKILL.md under version control. Run the audit after source systems refresh and before the weekly decision meeting. Give it read access first. Make the registry and prior report available on every run. Require an explicit approval step for destructive or externally visible changes.

The first dependency list should be short and executable

The agent will not have everything it needs on day one. That is normal. Ask it to rank missing access by decision value ÷ setup effort.

A useful first-week request sounds like this:

Connect read-only GA4 and CRM reporting, export the scheduled-job list, and provide the last four weeks of usage logs. This will let the next run connect jobs to qualified leads, estimate cost per useful output, and find duplicate data pulls. Owner: Maria. Time needed: 10 minutes. Due: Thursday.

A bad request sounds like this:

Give me access to everything.

The goal is to create momentum without turning the audit into a six-month data warehouse project.

What most teams forget

Three things are usually missing from the first design:

1. A consumer test. If nobody reads, approves, publishes, assigns, or acts on an output, the task may be producing inventory instead of value.
2. A duplicate-run guard. A manual “run now” plus the normal schedule must not send the same client report twice.
3. A counterfactual. Ask what would happen if the job stopped for two weeks. If nobody can describe the loss, the job needs measurement before it needs more tokens.

Add a fourth for knowledge-heavy organizations: preserve learning. When a run discovers a missing instruction, a bad assumption, or a better source, feed that lesson into the shared method through Recursive Self-Improvement QA. Otherwise every scheduled job pays to relearn the same lesson.

Start with one honest weekly cycle

Do not wait for perfect attribution. Inventory the jobs. Connect the sources you already have. Label what you do not know. Make the first five-page decision pack. Then improve the evidence every week.

The point is not to have the most agents or the most scheduled tasks. The point is to build marketing, sales, and operations functions that reliably move the business—and to stop paying for work that only looks busy.

If you are starting from scratch, read How to Set Up AI Agents for Your Business, install the relevant files from the Skill Pack Library, and compare your department coverage with 18 Scheduled Tasks Every Agency Owner Should Build in Claude.