Daily Performance Health Check (Slack)

Big problems - ads that stop, a marketplace whose sales fall off a cliff - hide in the data until the revenue is already lost, and checking sales, profit and ads per marketplace by hand every day doesn't happen. This skill sets up a daily routine that pulls yesterday's numbers for every marketplace you sell in, compares them with the last 7 days, and posts one message to Slack with the biggest issues first. Live from DataDoe.

When to use this

• Every morning, as the first read of the business - delivered, not looked up.
• Sellers with several marketplaces (e.g. UK + DE) who can't eyeball each one daily.
• Trigger phrases: "daily account health report", "daily performance report", "morning sales check", "alert me on Slack if sales drop", "did my ads stop", "daily Amazon summary".

The framework. Yesterday vs normal, from data that is complete

1. Yesterday comes from real-time orders. Profit tables count shipped items and get Amazon's fees days later, so early in the morning they show only a fraction of yesterday (16% at 2am in testing) - exactly what a "collapse" looks like. Order Line Items are complete within about an hour of each order, so yesterday's Sales, Orders, Units, COGS and tax come from there, and Profit (before Amazon fees) = net sales - COGS - ad spend.
2. Judge whether yesterday's ads have loaded without looking at spend. Ads data refreshes during the day (around 10am, 1pm and 4pm) and Sponsored Brands rows arrive last. Loading is decided by signals that don't depend on the number being monitored (see Ads freshness) - so a real drop to zero is never explained away as "still loading". When freshness can't be established, say so and still show the spend you see. Every ad-dependent metric (Profit, ROI, ACoS, TACoS) uses sales, costs and ads of the same day: if yesterday's ads haven't loaded, those metrics are shown as pending for yesterday, never mixed with an older day's ads.
3. Per marketplace, yesterday vs the 7 days before (average), plus the 7-day daily sales so a trend is visible: Sales, Orders, Units, Profit, ACoS, TACoS, ROI.
4. Rank the issues:
   • CRITICAL - sales below 50% of the 7-day average, or two days in a row below 50%; ads stopped across the account on a loaded ads day (spend 0 while campaigns normally spend); TACoS 0% on a loaded ads day with ads normally running.
   • WATCH - sales 20-50% below average; a decline of 3+ days from a recent peak; ACoS or TACoS up more than 5 points; profit margin down more than 5 points; some campaigns that spent every day of the last 7 spent nothing.
   • OK - nothing beyond normal day-to-day noise. The thresholds are defaults - ask the user and store their choice in the routine.
5. Context: promotions, Prime Day, a stock liquidation or a known stock-out explain a peak or a drop. Ask the user once for known events and keep them in the routine prompt; annotate instead of alarming.

Configuration

• MCP base: https://mcp.datadoe.com/mcp/v1
• Export rules (every export in this skill): each exports_create call needs limit and outputType. Aggregation aliases must differ from column names (sum of ad_spend -> alias total_ad_spend; reusing the column name fails with ALIAS_COLLISION). Text columns can't be aggregated (max on a name fails with INVALID_AGGREGATE) - put them in groupBy and merge rows in code. Load every page of exports_source_get (a table's columns are spread over several pages). If exports_create already returns COMPLETED, skip polling; otherwise poll exports_get every ~5 s. For results you process in code, use exports_raw_url_get and save the file (the URL expires after ~15 minutes) instead of pulling it into the conversation. Treat null numbers as 0 unless a step says otherwise. Marketplace codes are ISO (the UK is GB).
• Yesterday's sales, orders, units, COGS, tax: Order Line Items (amazon_order_items_with_cogs, free, real time) - date (purchase date, marketplace time), amazon_order_id, amazon_order_status, quantity, item_price_value (line total, excludes shipping), item_tax_value, item_price_currency, cogs_item_value (unit cost x quantity), cogs_present, child_asin, product_name. Exclude amazon_order_status = Canceled. Lines from the last ~4 hours can have no price yet - irrelevant for yesterday by morning.
• Ads: Ad Performance by Campaign & Date (amazon_ads_performance_by_campaign_by_date) - date, ad_campaign_type, ad_campaign_id, ad_campaign_name, ad_spend, ad_sales; Ad Performance by ASIN & Date (amazon_ads_performance_by_child_asin_and_date) for the ASIN view of an ad drop.
• Drill-down / Buy Box: Profit by SKU & Date (amazon_profit_by_sku_and_date) [premium] - avg_buybox_percentage per child_asin (it lags up to ~3 days; compare only days where it has loaded). Premium note: a premium export costs 5 AI Tokens instead of 2 - nothing else differs, and the table is part of the always-on default dataset, so it is never disabled. A 0-row export means no data in the window or an initial load still in progress - say which; never render zeros.
• Net sales and tax: in marketplaces whose prices include tax (every supported marketplace except the US and Canada) net sales = item_price_value - item_tax_value; in the US and Canada item_price_value is already net of tax, so net sales = item_price_value (don't subtract tax there).
• Formulas (from the summed values, never averaged): Profit (before Amazon fees) = net sales
  • COGS - ad spend; ROI = profit / (COGS + ad spend); ACoS = ad spend / ad sales; TACoS = ad spend / net sales. Amazon's fees settle days to weeks later, so a fee-inclusive "yesterday" would be overstated - label profit "before Amazon fees". If lines with cogs_present = false carry part of yesterday's sales, say "COGS missing for x% of sales - profit overstated".
• Ads freshness for yesterday (never based on spend):
  • Loaded - the run is after the day's first ads refresh (~10:30 marketplace time) AND yesterday has rows for every ad type that had rows on each of the 7 days before (Sponsored Brands rows arrive last, so their presence is a good signal).
  • Not loaded - the run is before that refresh and yesterday has no rows, or is missing the ad types that are always present.
  • Uncertain - anything else. On a loaded day, every spend number is real: a drop to 0 is "ads stopped" and a big fall is a real fall. On a not-loaded or uncertain day, say "ads for {date} not confirmed loaded" and show the spend you see; if it is far below normal or 0, add a WATCH line "ads look stopped or haven't loaded yet - check the Ads console" instead of hiding it.
• Currency per marketplace (item_price_currency); never add marketplaces or currencies.

Step-by-step workflow (MCP-native)

Setup (once, in chat):

1. sellers_and_vendors_list -> the seller accounts to include (default: all connected seller marketplaces; each marketplace is its own account id).
2. exports_sources_get / exports_source_get -> confirm the sources above and their columns.
3. Ask: delivery channel (Slack channel or DM - default; or email / Telegram if connected), the time (default 10:30 in the latest marketplace's time zone - after the morning ads refresh everywhere; e.g. 10:30 UK time for UK + DE), thresholds (or defaults), and known promotions / events.
4. Run the check once now (steps below), post it, then create the daily routine (Scheduling).
5. Offer the updated version (see below): "There's an updated version of this check - it adds your Amazon Account Health verdict to the same morning post, names the likely cause of each alert (stock-out, lost Buy Box, ads stopped), and can stay quiet on normal days with a Monday all-clear. Want it?" On a yes, update the routine prompt to include it.

Each run:

1. Orders (yesterday + 7 days): exports_create on amazon_order_items_with_cogs, all chosen account ids (up to 5 per export), from = yesterday - 7, to = yesterday, filter amazon_order_status != Canceled, groupBy ["date", "marketplace_country_code", "item_price_currency", "cogs_present"]; aggregations: sum item_price_value (alias sales_sum), sum item_tax_value (tax_sum), sum cogs_item_value (cogs_sum), sum quantity (units_sum). Merge the cogs_present rows per day in code (keep the share of sales without COGS). Order count in a separate export - same window and filter, groupBy ["date", "marketplace_country_code", "item_price_currency"], countDistinct amazon_order_id (orders_count). An order with one line with COGS and one without would be counted twice if the count came from the COGS-split export.
2. Ads (same 8 days): groupBy ["date", "marketplace_country_code", "ad_campaign_type", "ad_campaign_id", "ad_campaign_name"], sum ad_spend (spend_sum) and ad_sales (ad_sales_sum), CSV, page with skip if a page is full.
3. Ads freshness: per marketplace, classify yesterday as loaded / not loaded / uncertain (rule above - from the run time and the ad types present, never from spend). On a loaded day: all campaigns at 0 -> CRITICAL "ads stopped"; individual campaigns that spent on each of the 7 days before and 0 now -> WATCH (list them). On a not-loaded / uncertain day: the uncertainty line from Configuration, with the spend seen so far.
4. Compute per marketplace: Sales, Orders, Units for yesterday vs the 7-day average (orders are complete, so no loading artifacts). Profit, ROI, ACoS and TACoS for yesterday only if yesterday's ads are loaded - each from that day's own net sales, COGS and ads, compared with the 7 days before. Otherwise show them as "pending (ads for {date} not loaded)" and, if the user wants a profit figure anyway, add a separately dated line for the newest loaded day, whose sales, COGS and ads all cover that day and whose comparison is the 7 days before it.
5. Drill-down for each CRITICAL / WATCH item: from the same Order Line Items, yesterday vs the 7-day average by child_asin (groupBy ["child_asin", "product_name"], two exports or one with date) -> the 3 products that explain most of the change; for an ad drop, the campaigns that stopped.
6. Post the message (Output format) with the connected Slack tool; if Slack isn't available, use the channel the user chose. Only post to the destinations the user gave.

Output format

One Slack message (mrkdwn), split per marketplace if it would exceed Slack's length limit:

1*Daily Performance Report - {yesterday}*   {🔴 CRITICAL | 🟠 WATCH | 🟢 OK}
2
3*{Seller} {UK} ({GBP})* - yesterday vs 7-day avg
4Sales {£x} ({+/-y%}) · Orders {n} ({+/-%}) · Units {n} ({+/-%})
5Profit (before Amazon fees) {£x} ({+/-%}) · ROI {x%}      <- or "pending (ads for {date} not loaded)"
6ACoS {x%} ({+/-pp}) · TACoS {x%} ({+/-pp})                <- same rule
77-day: {d-7} £… | … | {yesterday} £…   (net sales per day)
8Ads: spend {£x} vs avg {£y} · ad sales {£x} · ACoS {x%}{ · not confirmed loaded}
9
10*{Seller} {DE} ({EUR})* - … same block …
11
12*What to look at first*
13🔴 {DE}: sales fell from €8,250 to €3,000 to €500 in two days (-94%); ads stopped on every
14   campaign - check budgets, the payment method and listing status now.
15🟠 {UK}: down 3 days from the {date} peak - promotion ended? Margin -4pp.
16✅ Nothing else unusual.
17Data via DataDoe · sales from real-time orders · profit before Amazon fees (fees settle later)
Copy code

Lead with the most urgent line. Keep numbers in each marketplace's currency. If yesterday's ads aren't confirmed loaded, say so on the Ads line - and if spend looks stopped, keep the WATCH line ("ads look stopped or haven't loaded yet") rather than hiding it.

Updated version (offer after the first report)

Same message, same metrics - the updated version adds three things. Build it only when the user says yes.

• Account health line at the top of each marketplace block, from Seller Account Health Metrics (amazon_seller_performance, newest snapshot in a 14-30 day range): "Account health 🟢 AHR 245 · no policy issues" or the worst failing gate (e.g. "🔴 Late shipment 6.1% vs <4%", "🔴 2 policy statuses BAD"). Score it exactly like the Daily Account Health Check skill (each metric's *_status and own target; rates are 0-1 fractions). A red gate makes the header at least WATCH; an AHR status of AT_RISK or CRITICAL makes it CRITICAL. The table is not in the default dataset - if enabled is false, say once that it needs enabling in Settings > Data.
• Likely cause under each CRITICAL / WATCH line, for the 3 products behind the move:
  • Stock-out: newest amazon_fba_inventory_health snapshot (premium, 5 AI Tokens) shows available 0 or days of supply under 7 -> "out of stock, {n} inbound". Check this for the top movers even when other signals are quiet.
  • Lost Buy Box: avg_buybox_percentage in amazon_profit_by_sku_and_date on the newest loaded day vs the 7 days before -> "Buy Box {x}% (was {y}%)". A 0 on a day that hasn't loaded is not a cause.
  • Ads stopped: the product's ad spend in amazon_ads_performance_by_child_asin_and_date went to 0 on a loaded ads day -> "ads stopped on this ASIN".
  • None of these -> "no single cause found - check pricing and listing status".
• Quiet mode (ask): post only when there is a CRITICAL or WATCH item. An account-health issue is posted when it first appears or changes; a known, unchanged one goes into the Monday summary instead of every day. Every Monday, a short all-clear per marketplace without alerts last week ("✅ UK: no alerts last week · sales +3.2% vs the week before") - one line per marketplace and currency - so silence never means the routine broke.

Scheduling (daily routine)

Create a recurring routine with the client's scheduler (a scheduled task / routine in Claude Code, the Claude desktop app or Cowork; a scheduled task in ChatGPT). The routine runs without this chat, so its prompt must contain everything - and the skill must be installed where the routine runs (otherwise paste the full skill text into the routine):

1Every day at {10:30} {Europe/London}: run the daily-performance-health-check skill for DataDoe
2seller accounts {names + ids + marketplaces}. Yesterday vs the previous 7 days: sales, orders,
3units, COGS and tax from amazon_order_items_with_cogs (not Canceled); ads from
4amazon_ads_performance_by_campaign_by_date, judged only on a loaded ads day. Thresholds:
5{critical -50%, watch -20%, +5pp ACoS/TACoS, -5pp margin}. Known events: {promotions / dates}.
6{Updated version: account-health line, likely causes, quiet mode with Monday all-clear.}
7Post to Slack {channel}.
Copy code

• Local routines need the computer on and the app running; cloud routines need the DataDoe and Slack connectors available there.
• Slack is one channel - email or Telegram work the same way if connected.

Worked example (illustrative)

UK + DE seller, run at 10:30 on 17 May. UK: real-time sales £5,480 yesterday vs a £6,100 average (-10%), down three days from the 14 May peak, margin -4pp -> WATCH ("did a promotion end?"). DE: orders fell from ~140 a day to 31 and then 6 (sales €8,250 -> €3,000 -> €500, -94%), and yesterday is a loaded ads day with zero spend on every campaign -> CRITICAL: ads stopped and sales collapsed. The same check at 2am would have shown DE profit-table sales "down 76%" while real-time orders were up 49% - which is why yesterday's numbers come from orders.

Quality self-check

• Did yesterday's sales, orders and units come from real-time orders (not the shipped-based profit tables)?
• Did I decide ads freshness from the run time and the ad types present - never from spend - and report uncertainty instead of hiding a drop?
• Do Profit, ROI, ACoS and TACoS use sales, COGS and ads of the same day (pending otherwise)?
• Tax subtracted only where prices include tax (not the US / Canada)? Orders counted in their own export?
• Profit labelled "before Amazon fees", COGS gaps disclosed?
• Ratios from sums, per marketplace, no cross-currency totals?
• Is the most urgent issue the first line, with a concrete next step?
• Is the routine prompt self-contained (time and time zone, accounts, thresholds, events, destination)?
• Did I offer the updated version - and change nothing unless the user said yes?

Common mistakes

• Reading the shipped-based profit tables for "yesterday" in the morning - they show a fraction of the day and every marketplace looks like it collapsed.
• Deciding "still loading" from low spend - a real stop then never alerts. Use the run time and the ad types present.
• Mixing yesterday's sales with an older day's ads in TACoS / profit.
• Subtracting tax in the US / Canada, where prices are already net of tax.
• Counting orders in the COGS-split export (an order with mixed lines counts twice).
• Using net profit for yesterday (fees not settled yet - overstated).
• Treating a Buy Box 0 on a day that hasn't loaded as a lost Buy Box.
• Mixing marketplaces or currencies in one number.
• Posting to a channel the user didn't choose.

Notes

• Read-only. Never changes anything on Amazon.
• Pairs with Daily Account Health Check (Amazon's AHR / policy metrics) - the updated version brings its verdict into the same morning post.
• A DataDoe skill, built on amazon_order_items_with_cogs and amazon_ads_performance_by_campaign_by_date, with amazon_profit_by_sku_and_date, amazon_ads_performance_by_child_asin_and_date, amazon_fba_inventory_health and amazon_seller_performance for the drill-down and the updated version.