AI Prompts for Support Reporting and Weekly Reviews: 15 That Pull Real Numbers (2026)

The short version: these fifteen prompts turn support numbers into decisions. Not formatting: analysis. Which metric actually changed, what drove it, whether it is a trend or a bad Tuesday, and what to do about it on Monday morning. The first eleven work by copy-paste in ChatGPT, Claude, or Gemini: you bring the numbers, the AI brings the reading. The last four are different in kind: connected to an inbox through MCP, they pull the real response times, volumes, and conversations themselves, which makes this page the clearest demonstration in the library of what a connected assistant can do that a chat window cannot.

Where this page sits: the triage prompts and shared inbox prompts each carry one Friday-briefing prompt for the week's status. This page is what comes after the briefing: the analysis, the diagnosis, and the reports people upstream actually read.

The weekly review (prompts 1-4)

The weekly review fails two ways: it becomes a ritual nobody reads, or it becomes a number dump nobody interprets. These four force an interpretation.

1. The what-changed reading. The only weekly question that matters, asked directly.

Works in Claude, ChatGPT, or Gemini

Here are this week's support numbers next to last week's: [paste]. Which changes are material and which are noise? For each material one, give the most likely driver based on what else moved with it, and say what you would check to confirm.

Run this directly from Claude or ChatGPT with Drag→

Copy

2. The outlier autopsy. Averages hide the story; the worst three conversations usually tell it.

Works in Claude, ChatGPT, or Gemini

Here are our three slowest or messiest conversations this week: [paste summaries or threads]. What do they have in common: question type, handoffs, missing information, time of day? Name the pattern, not just the incidents.

Run this directly from Claude or ChatGPT with Drag→

Copy

3. The trend caller. Four data points are not a trend, but they might be the start of one.

Works in Claude, ChatGPT, or Gemini

Here are the last four weeks of our core numbers: [paste]. For each metric, call it: improving, worsening, flat, or too noisy to say. For anything worsening, say how many more weeks of data would make it a trend worth acting on, and what early action costs little if you are wrong.

Run this directly from Claude or ChatGPT with Drag→

Copy

4. The Monday plan. The review is only worth doing if next week starts differently.

Works in Claude, ChatGPT, or Gemini

Based on this week's numbers and the issues behind them [paste], write Monday's plan: the one process change to try this week, who it affects, and the single number that will show by Friday whether it worked.

Run this directly from Claude or ChatGPT with Drag→

Copy

Diagnosing the numbers (prompts 5-8)

A slow reply time is a symptom. These prompts find the disease.

5. The reply-time decomposition. "We are slow" is not a diagnosis. Where the time goes is.

Works in Claude, ChatGPT, or Gemini

Our first reply time worsened. Here is what I can tell you about the pipeline: when messages arrive, when they get assigned, when drafting starts, when replies send [paste what you have]. Split the delay across those stages, name the biggest one, and give one fix per stage that would cut it.

Run this directly from Claude or ChatGPT with Drag→

Copy

6. The volume forecaster. Next week's staffing question, answered from arrival patterns instead of vibes.

Works in Claude, ChatGPT, or Gemini

Here are our daily incoming volumes for the past month, plus anything unusual coming up (launch, billing day, holiday): [paste]. Forecast next week by day, flag the day most likely to overload us, and say what coverage that day needs.

Run this directly from Claude or ChatGPT with Drag→

Copy

7. The repeat-driver finder. The fastest way to shrink a queue is to stop a question from being asked.

Works in Claude, ChatGPT, or Gemini

Here are this week's conversation subjects or summaries: [paste]. Group them by underlying question, rank groups by volume, and mark which ones could be prevented by a docs page, a product fix, or a clearer email we send earlier. The top preventable group is next week's project.

Run this directly from Claude or ChatGPT with Drag→

Copy

8. The commitment-risk scan. The report that keeps promises from becoming apologies.

Works in Claude, ChatGPT, or Gemini

Here are our open conversations with ages and last-message summaries: [paste]. Which ones contain a promise we made (date, refund, callback) that is at risk of being missed? Rank by damage if missed, and draft the one-line internal nudge for each owner.

Run this directly from Claude or ChatGPT with Drag→

Copy

Reports people actually read (prompts 9-11)

The audience for support numbers is rarely the support team. These three translate for the people upstream.

9. The exec paragraph. Support health for someone with thirty seconds and a decision to make.

Works in Claude, ChatGPT, or Gemini

Turn this week's support numbers and issues into one paragraph for a founder: [paste]. Lead with the single most decision-relevant fact, keep two supporting numbers, cut everything else. End with the one thing you need from them, or say explicitly that you need nothing.

Run this directly from Claude or ChatGPT with Drag→

Copy

10. The board line. Two sentences, quarterly trend, no adjectives doing the work of numbers.

Works in Claude, ChatGPT, or Gemini

From these three months of support metrics [paste], write the two-sentence support update for a board deck: the trend, the driver, and the number that proves it. If the story is bad, say it plainly with the fix underway; a board smells a buried problem faster than a stated one.

Run this directly from Claude or ChatGPT with Drag→

Copy

11. The release correlator. Whether the product change moved the support queue, argued with dates.

Works in Claude, ChatGPT, or Gemini

We shipped [change] on [date]. Here are daily volumes and question types for two weeks either side: [paste]. Did the release change what people contact us about? Separate volume change from mix change, and say plainly if the data cannot answer the question.

Run this directly from Claude or ChatGPT with Drag→

Copy

The prompts that pull their own numbers (12-15)

Everything above depends on what you paste, which means the analysis is only as good as your export. Connected through MCP, the assistant queries the inbox's real analytics: response times, daily activity, closed conversations. Nothing is remembered, estimated, or copied from a dashboard. The connection guide covers the setup in about ten minutes.

12. The self-writing weekly report. The full Friday review, from live data to finished paragraphs.

Works in Claude, ChatGPT, or Gemini

Write this week's support report from our inbox: volume in and out, first reply time against last week, resolution counts, the three longest conversations with a line on why, and anything still open past 48 hours. Then add the analysis: what changed, what drove it, and the one thing to fix next week.

Runs on a connected inbox→

Copy

Reply, using Drag's MCP tools

Week in review: 214 conversations in, 209 closed. Median first reply 2h 41m, up from 1h 58m last week; the slip is concentrated on Tuesday and Wednesday, and it is pickup time, not drafting time. Longest thread: a payment-webhook issue, four days, twice waiting on the vendor. Still open past 48 hours: three, all assigned. The one fix for next week: Tuesday's queue sat unassigned for three hours; a morning assignment rotation removes most of the regression.

What happens: get_response_times and get_daily_activity supply the week's real numbers, get_closed_activity shows what was resolved, and the long-runners are read, not summarized from memory. The report cites its own receipts.

13. The reply-time drilldown. Prompt 5's diagnosis, run on live data instead of your reconstruction.

Works in Claude, ChatGPT, or Gemini

Break down our first reply time this week: by day, and by conversation. Find where the delay actually lives: which days, which kinds of question, and whether the slow ones were slow to be picked up or slow to be answered once picked up.

Runs on a connected inbox→

Copy

What happens: the assistant pulls per-day response times and reads the slow threads themselves, so the pickup-versus-drafting split comes from timestamps rather than recollection. The answer usually surprises the team lead.

14. The month-over-month review. The trend question, answered with every week's real numbers.

Works in Claude, ChatGPT, or Gemini

Compare this month's support performance to last month's: volumes, reply times, resolutions. Call each metric as improving, worsening, or flat, show the numbers behind each call, and name the single most important change to investigate.

Runs on a connected inbox→

Copy

What happens: get_avg_response_time and the activity tools cover both periods, so the comparison is complete rather than sampled. Trend calls come with the arithmetic attached.

15. The daily anomaly watch. The thirty-second morning check that catches the bad day early.

Works in Claude, ChatGPT, or Gemini

Compare today so far against our usual pattern for this weekday: volume, response times, anything aging unusually. If everything is normal, say so in one line. If something is off, name it, quantify it, and point at the threads involved.

Runs on a connected inbox→

Copy

What happens: the assistant checks live activity against the recent baseline and answers in one line when the answer is boring, which is the mark of a monitoring habit that survives. The one-line "all normal" is the feature, not the fallback.

Where the prompts stop and the system starts

Reporting prompts work exactly as well as the data underneath them, and pasted data decays: it is partial, stale by Friday, and blind to whatever your export missed. That is the honest limit of prompts 1 to 11, and it is why 12 to 15 exist. For disclosure: they run against Drag, our shared inbox for Gmail, which tracks response times and team activity natively and exposes them to any assistant through its MCP server on the Pro plan. Drag runs from $12 a seat with Drag AI from $18. Everything in the first eleven prompts needs no Drag account at all.

FAQ

Can ChatGPT analyze customer support metrics?

Yes, within the limits of what you paste. Given clean numbers, it is good at exactly the analysis teams skip: separating material changes from noise, decomposing a slow reply time into stages, and forcing a Monday decision from a Friday review. Connected to the inbox through an MCP server, it can pull those numbers itself instead of working from your export.

What should a weekly support report include?

Four things: what changed against last week and why, the outliers and their shared pattern, the promise-at-risk list, and one process change for next week with the number that will prove it worked. Volume and reply-time totals belong in the report; a report that is only totals is a ritual, not a review.

How do I report support metrics to a founder or board?

Lead with the decision-relevant fact, not the dashboard. A founder needs one paragraph: the trend, its driver, two numbers of evidence, and the ask. A board needs two sentences and no adjectives standing in for figures. Bad news stated plainly with the fix underway reads better than good news that smells curated.

The rest of the prompt library

This page is one of seven in the Drag prompt library. The triage prompts clear the queue these reports measure, the support QA prompts review the replies behind the numbers, the shared inbox prompts run the team day-to-day, and the customer service prompts handle the tickets that become this page's outliers.