The point where the Command Center stops being a chat window and becomes infrastructure. Three automations a 7-8-9 figure operator actually wants, the exact prompts, the scheduling mechanic for each, and the limit that breaks every "set it and forget it" promise.
A daily report you have to remember to ask for is not a report. It's a chore. The whole point of wiring Claude into your data is that the brief lands in Slack before you open your laptop, the spend-spike pings you mid-meeting, and the Monday exec summary writes itself off numbers you already trust.
That part works. What doesn't work is the marketing line that any of this runs forever, untouched, on its own. The scheduling layer has hard edges, and most "automation" content quietly skips them. I won't.
Three automations below. Then the part nobody puts in the sales deck: what survives unattended, and what dies the moment you close your laptop.
A model can do arithmetic wrong on a large CSV, hallucinate a column, or run silently on stale data. In a chat you catch that, because you're reading the answer. In an automation, nobody reads until the number is already in Slack in front of your team.
So every automation here ships with a verification step baked into the prompt and a guardrail around the schedule. Treat the Command Center as a fast, sharp analyst who occasionally fat-fingers a number, never as a system of record. Your system of record is still Shopify, Stripe, your bank, and your warehouse. Running an automation unattended doesn't change that. It raises the stakes on it.
One thing before the prompts. If you run these on live connectors rather than CSVs, most of those connectors carry their own bill on top of your Claude seat, and the size of that bill depends on which source you're pulling. The first-party MCPs are the cheap path: Meta Ads (official MCP at mcp.facebook.com/ads, write-capable, on Marketing API v25.0), Google Ads (official, read-only, 3 tools), GA4 (official, read-only), Klaviyo (official remote, 16 tools, OAuth on connect, no API key setup), Stripe (official, ~25 tools, remote OAuth at mcp.stripe.com), and BigQuery (official remote MCP). Source: gomarble.ai/ai-tools/meta-ads-ai-connectors, developers.google.com/google-ads/api/docs/developer-toolkit/mcp-server, developers.google.com/analytics/devguides/MCP, developers.klaviyo.com/en/docs/klaviyo_mcp_server, docs.stripe.com/mcp, docs.cloud.google.com/bigquery/docs/use-bigquery-mcp (all captured June 2026, verify). Those are free or near-free to connect, but they still bill API usage per query on the Claude side.
Where there's no first-party server, you're either calling the API yourself or paying third-party middleware: Windsor.ai, Porter Metrics, Supermetrics and similar each carry their own subscription, and exact pricing varies by vendor and data volume and isn't consistently published. Source: getpassionfruit.com, windsor.ai, portermetrics.com, supermetrics.com (captured June 2026, verify; treat per-brand connector cost as unverifiable, do not quote a fixed figure). Two sources have no clean Claude-native connector at all: Shopify live store data (the official Shopify MCP is docs/schema only, not your orders, so you pull the Admin GraphQL API yourself) and your bank cash position (Plaid/Mercury/CSV, no first-party path). Source: shopify.dev/docs/apps/build/devmcp, weaverse.io/blogs/shopify-ai-toolkit-dev-mcp-hydrogen-2026 (captured June 2026, verify). Nothing here is "free to run" once it's pulling live data. Say that to your team before you build it.
The 9 AM brief. Yesterday's net revenue, total ad spend, blended MER, order count, new-customer count, and returning-revenue share, each compared to the same weekday last week, posted to a channel as five tight lines.
Why same-weekday and not "yesterday vs the day before": DTC revenue has a weekly shape. Monday against Sunday tells you nothing except that Sunday is Sunday. Monday vs last Monday tells you whether the business moved.
One thing that separates this from a naive digest, and the reason an operator should care: ad platforms restate recent days for 24-72 hours as conversions attribute late. Source: cc-architecture build doc, captured June 2026. So a digest that reads "yesterday" off platform-claimed numbers is reading a figure that isn't final. The prompt below pins revenue to Shopify (the source of truth, not the ad platform, which lets each of Meta, Google and Klaviyo claim the same order and triple-counts if you sum them) and forces the model to print the data's max date so a late-restated or lagging feed is visible on every run.
Every morning, pull yesterday's numbers: net revenue, total ad spend, blended MER,
order count, new-customer count, and % of revenue from returning customers.
Definitions to use exactly:
- Net revenue = Shopify gross sales minus refunds and discounts. This is the revenue
source of truth. Do NOT use ad-platform-claimed revenue. Tell me which column you used.
- Total ad spend = sum of every paid channel's spend for the day.
- Blended MER = net revenue / total ad spend.
- New customer = a customer whose first-ever order falls in this window.
Compare each metric to the SAME WEEKDAY last week, not to the prior day, to control for
weekly seasonality.
Write a 5-line Slack digest: each headline number, its week-over-week delta as a % and an
arrow, and one sentence flagging anything notable. Post it to #daily-numbers.
VERIFY before posting: state the max date actually present in the data so I know it
includes yesterday and isn't stale or mid-restatement. If any metric could not be computed
because data was missing or a date was blank, say so in the digest in plain words. Do not
silently skip a metric and do not invent a number to fill a gap.
The verify clause is the whole point. The failure mode of a daily digest isn't a wrong number once. It's a digest quietly reporting on two-day-old data for a week because a connector lagged and nobody noticed. Forcing the model to print the max date present makes staleness visible on every single run.
A daily, on-the-hour job with no machine dependency, which makes it the textbook fit for a cloud routine. Set it once, it runs on Anthropic's infrastructure, your laptop can be shut. Source: code.claude.com/docs/en/scheduled-tasks (captured June 2026, verify).