# Weekly Report Automation Recipe

Turn the Monday export-and-paste routine into a review job. The assistant assembles the numbers and drafts the narrative. You verify and send.

**Works with:** ChatGPT, Claude, Gemini, or any capable AI assistant that accepts pasted text or file uploads. Nothing here depends on one vendor.

**What you get:** a one-time setup, a weekly routine, two copy-paste prompts, and a sanity-check list.

## Prerequisites

- An AI assistant that accepts file uploads or long pasted text
- Export access to every system your report covers (CRM, ad platforms, accounting, spreadsheets)
- One past report you were happy with
- Last week's report on hand each time you run this

## One-time setup

**1. Write the report spec.** List every metric in the report. For each one: the source system, the exact export or saved view that produces it, and what it's compared against (last week, target, same week last year). If a metric has no reliable source, decide now whether it belongs in the report at all.

**2. Save the export views.** In each source system, build and save a filter or view so the weekly export is the same two clicks every time — same columns, same date logic. Inconsistent exports are the top reason this recipe produces garbage.

**3. Extract your template.** Paste your best past report into the assistant with this prompt, and save what comes back:

```
Below the line is a past weekly report I was happy with.

Turn it into a reusable template:
1. Replace every specific number, date, and name with a placeholder in {curly braces}.
2. Keep the section order, headings, and tone exactly as they are.
3. Under each heading, add a one-line note on what the section covers and how long it should run.

Return only the template.

---
[PASTE YOUR PAST REPORT HERE]
```

## Weekly run

**4. Export.** Pull each saved view. Don't edit the files by hand.

**5. Assemble.** Start a fresh chat. Upload or paste this week's exports, last week's report, and your template. Then run:

```
Assemble my weekly report for {DATE RANGE}.

You have: this week's exports (listed below), last week's report, and my template.

Exports:
- {SOURCE 1 — e.g. CRM pipeline: new leads, deals won, deal value}
- {SOURCE 2 — e.g. ad platform: spend, conversions by campaign}
- {SOURCE 3 — e.g. accounting: invoiced, collected, outstanding}

Rules — follow these exactly:
1. Use only numbers that appear in the exports. Never estimate or fill a gap from memory.
2. If a template metric is missing from the exports, write [MISSING: metric] in its place. Never skip it silently.
3. If two exports disagree on the same figure, show both values with [CHECK: sources disagree].
4. Compare each metric to last week's report. Tag any move that looks out of line with the recent trend as [CHECK: verify] — do not invent an explanation for it.
5. Match the template's section order and tone. Short sentences. No hype.

Output:
A. The completed report.
B. A "Review before sending" list: every [MISSING] and [CHECK] flag, with one line on what to verify and where.
```

**6. Review.** Work through the "Review before sending" list first, then run the sanity checks below.

## Sanity check (do not skip)

- Spot-check three numbers at random against the source exports, including at least one the narrative calls out.
- Confirm every number in the narrative also appears in the data. A figure that exists only in prose is the classic hallucination.
- Confirm date ranges match across all sources. The most common silent error is one system exporting a different week.
- Resolve every [CHECK] and [MISSING] flag or cut the metric. Never ship a flag.
- Read the narrative once for claims the numbers don't support, and delete them.

If the same check fails two weeks running, fix the export view, not the prompt.

## When this outgrows DIY

This recipe holds while your sources are few, the exports are clean, and the report is internal. It stops being enough when sources multiply, when an export format changes and the report goes quietly wrong, or when the report goes to clients and one error costs trust. At that point the job isn't prompting — it's a monitored pipeline that pulls directly from your systems, checks totals against the source, and escalates when something is off. That's what we build and run at agentclaw: see https://agentclawhq.com/services, or book a free AI opportunity audit at https://agentclawhq.com/book and we'll map exactly what your report would take.