Data-Analysis Skills: Let Agents Own Reports & Retrospectives

Stop writing reports; start owning the report system

Weekly sales recaps, ad performance reviews, monthly retrospectives: most teams treat these as human chores. But the work has a fixed pattern. You pull data, compare it to a baseline, mark anomalies, write conclusions. That pattern can be handed to an agent.

The point is not to replace the analyst. It is to lock the analyst's judgment into a repeatable skill so every report uses the same definitions, the same output structure, and the same traceable logic.

In our own operations, this is how we run retrospectives inside 365Loopa. The queue to evidence to decision to log workflow is a natural home for a data-analysis skill: each report becomes evidence, not opinion.

What a data-analysis skill actually locks down

A weak AI report looks convincing and invents causal stories. A production-grade data-analysis skill removes that risk by fixing five things before the agent touches a single row.

The soul of a data-analysis skill is not calculation. It is consistent caliber plus traceable conclusions. If the agent cannot show which column, file, or query produced a number, that number is not usable for decisions.

Anatomy of a SKILL.md for reports and retrospectives

In 365Skill, a skill is a standard SKILL.md file with instructions, inputs, and evaluation criteria. A data-analysis skill should contain at least these sections.

• Trigger: When to run. Example: every Monday after 10:00, or when a new weekly sales export is uploaded.
• Data source: Exact file path, sheet name, table, or API endpoint. No fuzzy language.
• Fixed calibers: Metric definitions such as paid revenue, refunds, repeat customers, CAC, ROAS, and the comparison window.
• Output template: A table block, one chart block, a conclusion block, and an anomaly block.
• Conclusion grading rules: Every output line must be labeled Fact, Inference, or Recommendation.
• Traceability rule: Every fact must include its source pointer. Example format: File name, column, and row range or query name.
• Failure rules: What to do when data is missing, a metric is outside expected range, or a source changes format.

This is not engineering work. It is operating documentation. You are writing down the rules your best analyst already uses.

How to build your first data-analysis skill

Start with one recurring report. Do not try to automate all reporting at once.

1. Pick the most annoying weekly or monthly report. Choose one with stable inputs and a clear owner.
2. Lock the data source. Export one clean file or define one API call. Never let the agent search across multiple sources for the same metric.
3. Write calibers as if a new hire needs them. For revenue, state order status, refund treatment, currency, time zone, and cut-off time.
4. Define the baseline. Example: compare to the same 7-day window last week, not a rolling average that changes.
5. Create the output template. Require a table, a chart, conclusions, and anomaly markers. No narrative before the table.
6. Force conclusion grading. A line that says “conversion dropped because of the new landing page” must be marked Inference unless the source data proves causation.
7. Make traceability non-negotiable. If the agent outputs a fact without a source pointer, the skill has failed.
8. Run it in parallel for two weeks. Compare the agent report to the human report before switching ownership.

This checklist is the difference between a toy report generator and a skill you can trust for decisions.

Worked example: weekly cross-border sales retrospective

Imagine you run a small cross-border team selling through Shopify and running Meta ads. Every Monday, you want one retrospective.

The skill input

• Orders export from the store admin for the last 14 days.
• Ads export from the ad account for the same period.
• Currency conversion fixed at the weekly closing rate.

Fixed calibers

• Paid revenue means orders with financial status paid, excluding refunds and test orders.
• Comparison baseline is the previous 7-day calendar window, not the previous period of a campaign.
• ROAS means paid revenue divided by ad spend, not gross merchandise value.

Required output blocks

• Table: revenue, orders, average order value, ad spend, ROAS, repeat purchase rate.
• Chart: daily paid revenue versus baseline.
• Anomalies: any metric moving more than 20 percent versus baseline.
• Conclusions: Facts, then Inferences, then Recommendations.

Example output line

Fact: Paid revenue fell from 41,200 USD to 37,800 USD week over week. Source: orders export 2025-06-10, column paid_revenue_usd, rows 2 through 1,204.

Inference: The drop is concentrated in one product category because category-level revenue fell 31 percent while the rest stayed flat. Source: same export, category column grouped by week.

Recommendation: Review ad delivery and inventory for that category before changing overall budgets.

The agent did not invent a story. It surfaced an anomaly and kept the causal claim separate from the raw fact. That is the behavior you want.

This same structure powers our own retrospectives in 365Loopa. Evidence is queued, the skill produces the report, a human makes the decision, and the decision is logged. Next week, the baseline includes last week's reality.

What not to let the agent do

• Do not let the agent define metrics on the fly. Calibers must be read from the skill, not invented per report.
• Do not allow conclusions that mix facts and suggestions in one paragraph.
• Do not accept a report without source pointers. “Trends suggest” is not traceable.
• Do not let the agent silently fix missing data. Require an explicit data quality note.

If you cannot audit the report back to a raw file, you are reading hallucinated management consulting.

FAQ

Can data-analysis skills work with spreadsheets, dashboards, and APIs?

Yes. The skill should specify one source per metric. A spreadsheet export works well for weekly reports; an API works better when reports run daily or need near-real-time data. What matters is that the source is fixed and versioned.

How do I keep an agent from inventing numbers?

Force every output line to carry a source pointer back to a file, column, row range, query, or API response. Add a decision rule: if no source is available, the agent must say the data is missing and stop. Conclusion grading also helps because inferences are less dangerous when they are labeled as inferences.

Do I need to be technical to write a data-analysis SKILL.md?

No. The skill is mostly operating rules: metric definitions, time windows, output order, and failure behavior. If you can explain how you want the weekly report built, you can write the skill. The 365Skill library is designed for this kind of operator-led setup.

How does this connect to 365Loopa's evidence chain?

The evidence chain moves from queue to evidence to decision to log. A data-analysis skill produces the evidence. The report becomes a stable, traceable artifact that feeds the decision step. In 365Loopa, this removes the usual back-and-forth where everyone argues about which version of the data is correct.

If you want to stop rebuilding reports every week, start with one skill and one fixed caliber. If you need a faster entry point, try the free AI tools for product research, listing generation, and cold outreach, or book a free consult if you want help turning a recurring report into an agent-owned system.