Give AI your weekly sales report.

Give it the result and the next run.

“Summarize this spreadsheet” leaves a lot of the job with you. You still need to decide which numbers matter, notice a duplicate, reconcile the total, explain a missing owner, and work out how to repeat the process.

A stronger assignment makes those part of the deliverable. The agent owns a usable report, the evidence behind it, and the instructions that carry your decisions into the next run.

This follows Matt’s approach to briefing agents: describe the outcome, provide the real context, and set a clear standard for done. Do not spend the prompt dictating every operation the agent already knows how to choose.

A repeatable report remembers what the numbers mean.

That means the reporting brief matters as much as the prose summary. If “sales” quietly means something different next week, a consistent-looking chart can hide an inconsistent calculation.

The business definition comes before the total.

Our fictional Northstar case measures net completed sales in USD: count unique paid and refunded transactions, using their signed amounts. Exclude pending and cancelled transactions. Remove one copy of an identical duplicate. Keep a valid sale with a missing owner, but label that owner “Unassigned.”

Those rules are supplied business context. They are not universal accounting rules, and this metric is not profit, cash flow, or recognized revenue.

Without the definitions, an agent can produce a tidy report about a different number. Summing the raw original export gives $67,000. That includes a $1,800 duplicate, $7,000 cancelled transaction, and $3,500 pending transaction. Removing those exclusions gives $54,700.

What happened in the recorded rehearsal.

A coding agent received only the approved exercise inputs and a complete assignment. It wrote a reusable JavaScript analysis and HTML reports for the two source snapshots. The original source files were preserved. We then checked the saved results against a separate calculation.

The first verification caught a display issue: the report showed a minus sign on zero. The agent fixed it and reran the checks. The pack preserves the first generated output and the correction log; the numeric results did not change.

The second snapshot makes two explicit corrections: a $3,500 pending transaction becomes paid, and a $1,500 completed sale is assigned to Maya. It does not contain a new reporting period.

The paid-status correction adds $3,500 to the net total. Assigning the other sale changes the owner breakdown without changing the business total. Those are different kinds of changes, and the report needs to keep them separate.

The same calculation handled both snapshots. That is evidence for this small exercise, not a reliability claim for an arbitrary CRM export. A real deployment would also need to handle your actual schema, currencies, access rules, and exception policy.

The full outputs, exact assignment, source hashes, and run notes are in the pack. The reports were generated by a coding agent; they were not produced or recalculated inside Excel, Claude for Excel, or Cowork.

The assignment to adapt.

Replace the bracketed details and attach the source export and definitions. Choose a format your agent can actually produce and your team can use. For the recorded exercise below, the deliverables were HTML reports, JSON results, and a reusable script.

Take responsibility for turning [sales export] into the [reporting period] sales report for [audience]. The report should help us decide [business decision]. Use [metric definitions] and [previous approved report, if available]. I want the result, the exceptions that matter, and a reliable way to do this again next time.

Own the analysis and the finished report, not just a summary of the rows. Choose a useful structure. Tell me what deserves attention, what the evidence supports, and what remains uncertain. If a previous period, target, cost, or explanation is missing, don't invent one.

The bar for done: I can trace the headline to the source; the summary, tables, and charts agree; and all the source rows are accounted for. Use the metric definitions consistently. Flag conflicting records instead of silently picking one. Keep a valid transaction in the total even if a descriptive field is missing, and make the gap visible.

Preserve the original export. Use a simple repeatable calculation where appropriate. Challenge the main total with a separate calculation, resolve discrepancies, and inspect the saved output with the tools available. Tell me what you could not verify, including any native spreadsheet or formula checks.

Make routine choices yourself. Ask only about an ambiguity that changes the meaning of the report. Deliver the report in [usable format], a short update with the decision it supports, and a reusable reporting brief: the inputs, definitions, known exceptions, and checks for the next run.

Do not send the report, change source-system records, or schedule anything unless I have authorized it. A working report and a reusable process are the deliverables here.

Example brief adapted from Matt’s published approach. The pack also includes the exact assignment used for the coding-agent rehearsal.

A good report knows what it cannot explain.

Willow is the largest client by net completed sales in both snapshots. Its original $17,800 is 32.5% of the $54,700 total; after the correction it is 30.6% of $58,200. That is a supported concentration fact.

“Willow is growing” is not supported. Neither is “the sales strategy is working.” There is no prior period, target, margin, or explanation of customer behavior in these inputs.

The useful next step is to focus the owner’s attention on what the report exposes: concentration, the refund, the cancelled transaction, and any unresolved source problems. The agent can recommend what to investigate without inventing why it happened or committing a colleague to a deadline.

That distinction is part of the assignment. Ask for a position, and ask the agent to show where the evidence stops.

Leave the next run ready.

The handoff should preserve the decisions that made this report trustworthy: which input to use, the reporting period, what counts, how to treat duplicates and missing values, and which totals must reconcile.

The included runner reads the approved exercise inputs and produces the report and result files. Its instructions explain how to rerun it. An analyst or a capable coding agent can inspect the calculation instead of relying on an undocumented chat history.

Before adapting it to real work, use an approved copy of your export. Run the new and previous process side by side, inspect exceptions, and make sure the report still means the same thing. Keep personal and confidential data inside tools your organization permits.

A repeatable script is one piece of automation. We did not connect a CRM, schedule a job, email a report, or observe this process over multiple weeks. Those require separate implementation and checks.

For the next assignment, give the agent the previous reporting brief as context. Let it carry the work forward, and make it explain any definition it needs to change.

What we did, and what we didn’t.

This exercise was prepared with coding agents on September 2, 2026. It is teaching material, not a product benchmark or a report about Matt’s business.

• InputsThe existing fictional Northstar sales export and a JSON snapshot containing two explicit corrections to the same June 2026 period. Twenty source rows per snapshot.
• Actual workA separately started coding agent received the exercise inputs, wrote a JavaScript runner, and generated HTML reports and JSON results for both snapshots.
• ReviewThe coordinating coding agent checked the saved numbers with a separate calculation and preserved the source hashes. This was agent review, not human sign-off. The process is recorded with the files.
• Not performedNo native Excel editing or recalculation, product comparison, CRM connection, scheduled execution, elapsed-week trial, human sign-off, or measured time savings.

The download includes the exact exercise assignment, preserved outputs, source hashes, and verification notes. The prompt in this article is the broader assignment to adapt to your own environment.