Build a traceable expense summary from six CSV rows with Orkas

By ColunTina. AI-assisted material prepared on behalf of Orkas for its product-discovery campaign. All amounts, records and outputs in this example are fictional. This is a worked document example, not a recorded run inside Orkas or a claim about a customer's accounts.

A summary is easier to review when each total points back to its rows. A grand total on its own hides useful questions: was a refund retained, did an uncategorized item disappear, and did the writer turn an incomplete ledger into an approval decision?

This small example produces a Markdown summary and a short handoff from one CSV. Orkas is an open-source, local-first desktop AI workspace. Commander can coordinate specialist agents around project files; here OfficeWorker organizes the table and ContentWriter explains the result without inventing missing evidence.

1. Start with a small, explicit input

Save this illustrative file as expenses.csv:

id,category,amount,currency,note
E01,Hosting,80.00,USD,Monthly service
E02,Design,120.00,USD,Template purchase
E03,Travel,45.00,USD,Local trip
E04,Hosting,20.00,USD,Additional storage
E05,Design,-30.00,USD,Partial refund
E06,Unassigned,60.00,USD,Category needs owner review

Put the interpretation rules in summary-rules.md:

All six rows use USD. Do not apply an exchange rate.
A negative amount is a refund and reduces its category total.
Keep Unassigned as a separate category and include it in the grand total.
Retain the source IDs for every category. Do not remove a row without explaining why.
Receipts and bank statements are not supplied. The output is a summary of reported rows, not a reconciliation, reimbursement approval or tax classification.

In a real task, select only the necessary files and remove unrelated personal or payment information before providing them to an AI workflow. The illustrative notes above contain no actual supplier or employee identities.

2. Request the table and the explanation together

A possible instruction to Commander is:

Read expenses.csv and summary-rules.md. Use OfficeWorker to group the six rows by category, preserve the negative refund, list the IDs behind each total and calculate the grand total. Use ContentWriter to turn that same result into expense-summary.md with a short owner handoff. Show Unassigned separately. State which supporting evidence is absent. Do not approve payments, change source records, connect to an account or send the report anywhere. If you find a conflicting currency or unclear rule, identify it before choosing an interpretation.

These are specialist roles available in the Orkas catalogue, rather than proof that a particular installation has them enabled. Check the selected agents and file access in the workspace. This preparation task needs the supplied files; it does not need credentials to a banking or expense service.

3. Inspect the complete worked output

A consistent expense-summary.md would contain this table:

Category	Source IDs	Calculation in USD	Reported total in USD
Hosting	E01, E04	80.00 + 20.00	100.00
Design	E02, E05	120.00 - 30.00	90.00
Travel	E03	45.00	45.00
Unassigned	E06	60.00	60.00
All reported rows	E01-E06	100.00 + 90.00 + 45.00 + 60.00	295.00

The categorized subtotal is 235.00 USD. Another 60.00 USD remains unassigned. Together they equal 295.00 USD; keeping the open category visible avoids making the report look more complete than the input.

The accompanying handoff can be short and concrete:

Six reported rows are included. Hosting totals 100.00 USD; Design nets to 90.00 USD after the 30.00 USD refund; Travel totals 45.00 USD. E06 adds 60.00 USD to the grand total but still needs a category decision. The total of all reported rows is 295.00 USD. No receipts or bank statements were supplied, so these amounts have not been reconciled to external evidence. Next owner action: classify E06 and provide supporting records before any approval decision.

Review the IDs as well as the arithmetic. Each ID appears in exactly one category row, E05 stays negative, and the six input amounts sum to the same grand total. A readable paragraph is not a substitute for those checks.

4. Keep the method beside the result

Store the CSV, interpretation rules and summary together. Orkas Projects can keep instructions and memory for later work; selected Library documents can supply a relevant expense policy, and a reusable Skill can preserve the agreed report format. A policy that was not supplied should remain a named gap.

Local-first means working files and context stay on the device by default. Relevant content still goes to the chosen provider when a remote model is used. The core desktop app is free and MIT-licensed; model usage and optional hosted services can have separate costs.

For a first attempt, use a small non-sensitive file, require source IDs beside totals, and compare the output with the original rows. The official Orkas site explains the workspace and its workflows. Keep any later account access, edits, approvals or distribution within the actual task's authorized scope.