A Scenario Familiar to Many

Friday afternoon is the most dreaded time of the week for many indie developers and data leads on small teams. The SQL queries were written long ago, and the results are just sitting there—but the real time sink isn't pulling the data. It's turning a pile of cold, lifeless tables into something that reads like human language.

Here's my situation: I work on a product with a six-person team, and we track roughly a dozen metrics every week—new registrations, active retention, paid conversions, channel source distribution, and core feature usage frequency. Pulling the data itself takes ten minutes with ready-made SQL. But turning the results into weekly report text that leadership can read and colleagues actually want to read costs me two to three hours every single time.

Repeatedly checking whether numbers were copied correctly, whether the wording clearly explains the trends, whether the week-over-week and year-over-year figures are calculated right, whether the issues raised last week were followed up this week... These mechanical tasks happen to be exactly what AI excels at.

Breaking Down the Pain Points: Where Does the Time Actually Go

I once logged a complete manual weekly reporting workflow, and the time breakdown was roughly:

• Verifying numbers: 40 minutes. Copying query results into a document and confirming, one by one, that nothing was misaligned or missing.
• Calculating WoW/YoY changes: 30 minutes. Manually computing growth rates and digging up last week's report for baseline figures.
• Writing the narrative: 60 minutes. Translating tables into prose like "Metric X rose X% week-over-week, primarily driven by XX."
• Polishing and formatting: 30 minutes. Adjusting tone, adding conclusions, and appending next week's plan.

Of those three hours, the only part that truly required "human" judgment was that final bit of conclusive thinking. Everything before it was pure, well-defined repetitive labor.

That's exactly where AI fits in: AI doesn't make decisions for you, but it handles all the grunt work of moving from data to prose.

What AI Can Do for You

Concretely, if you hand SQL query results to AI, it can reliably do the following:

1. Automatically calculate week-over-week and year-over-year changes. Just give it this week's and last week's data together—no calculator needed.
2. Identify trends and anomalies. When a metric suddenly fluctuates beyond a threshold, it will flag it and, based on the dimension fields you provide, suggest possible causes.
3. Generate structured weekly report text. It produces a finished draft organized into sections you define (Overview, Highlights, Risks, Focus for Next Week), with consistent definitions and phrasing.
4. Maintain consistent terminology. By fixing metric definitions in the prompt, you avoid the confusion of calling it "active users" this week and "DAU" next week.
5. Multi-version output. A one-page summary for management, a detailed version for the team, an external version for investors—all generated in one pass.

One caveat: AI is not an analyst. It won't judge for you whether "this fluctuation is worth caring about." My approach is to let it produce a draft and flag questionable points, then spend five to ten minutes making judgments and corrections myself.

My Actual Workflow

Once the whole process was up and running, my fixed Friday routine became four steps:

Step 1: Run the SQL and export the results. Export the query results for the dozen-plus metrics as CSV, or copy them directly as Markdown tables. This step takes 5 minutes.

Step 2: Assemble the prompt. Use my fixed template (see next section), filling in this week's data, last week's data, and metric definitions. 3 minutes.

Step 3: Let AI generate the first draft. Generate in one pass; if a particular section isn't satisfactory, issue a follow-up instruction to rewrite just that part. 2 minutes.

Step 4: Human review. Focus on verifying the growth rates AI calculated and whether its attribution of anomalous fluctuations holds up, then add my own judgment. About 10 minutes.

That's about 20 minutes total, versus 180 minutes before—saving roughly 160 minutes per week, which at four weeks a month works out to over 10 hours—nearly a day and a half of work. For a small team, this essentially frees someone from the role of "report worker."

A Ready-to-Copy Prompt Template

Here's the template I've settled on. You can copy it directly and replace the placeholders:

你是一名数据分析师,请根据我提供的数据生成一份中文数据周报。

【角色与目标】
- 读者:团队负责人和非技术同事,避免术语堆砌
- 目标:让人30秒抓住重点,3分钟看完细节

【指标口径】(请严格遵守,不要自行更改定义)
- 活跃用户:当日登录且停留超过30秒的去重用户数
- 付费转化率:当周付费用户数 / 当周活跃用户数
- {在此补充你的指标定义}

【本周数据】
{粘贴本周查询结果,建议用表格}

【上周数据】(用于计算环比)
{粘贴上周查询结果}

【输出要求】
1. 一句话总览:本周整体表现(好/平/差)及核心原因
2. 关键指标表:数值 + 环比 + 简要解读
3. 亮点2-3条:涨幅或改善明显的指标,说明可能的驱动因素
4. 风险2-3条:跌幅或异常波动的指标,标注需要人工确认的疑点
5. 下周建议关注:基于数据提出的2条观察方向
6. 所有数字必须与原始数据完全一致,不确定的推断请标注"待确认"

【语气要求】
客观、简洁,不做没有数据支撑的归因。

The two most important design choices in this template: first, mandatory week-over-week output, which eliminates manual calculation; second, requiring a "pending confirmation" marker, which prevents AI from confidently fabricating attributions—I added that only after learning the hard way.

Before and After: It's Not Just About Saving Time

The most unexpected gain is the last one: because the cost dropped, I started doing granular analyses I used to think "weren't worth the time," like retention comparisons broken down by channel. The weekly report gradually transformed from a box-checking document into something the team actually discusses.

A Few Reminders

• De-identify your data. Before sending anything to AI, strip out sensitive fields like user IDs and phone numbers, keeping only aggregated metrics. If your data is highly sensitive, consider a self-hosted model or an enterprise API that doesn't participate in training.
• Model selection. This task doesn't demand much from a model—any mainstream conversational model can handle it. No need to chase the most expensive flagship; check the official pricing pages and estimate costs based on your usage volume.
• Keep the human in the loop. Let AI calculate the growth rates, but "should this anomaly be escalated" is always a human decision.

If you'd also like to reclaim those few hours every week, give this workflow a try. The whole solution relies on a stable LLM API—I'm using the API service at https://api.thistoken.ai/register. Sign up and you can get started right away—run the template once, and you'll feel the difference by next Friday.

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.