Use Claude Code to check an earnings comparison
A useful earnings task in Claude Code starts with two comparable source files and a narrow question. A persuasive summary alone is insufficient: you should be able to trace each amount to the input, recompute the change and identify what remains unknown.
Prepare the task, not a collection of PDFs
Use the Claude Code installation guide for the current package path and invocation. The official skills documentation describes the host mechanism; CoTrading compatibility still depends on the package’s recorded host test.
Create a research folder with previous and current reports, a question and an input manifest. Record identity, reporting window, currency, scale and whether a number is cumulative or point-in-time. Do not supply account credentials or a portfolio simply to compare statements.
The current 0.1.0 S01 package requires Python 3.12 for its bundled deterministic tool, just like the other A-batch methods S02/S04/S05. The tool uses only the standard library and no network for the text sample. Check runtime availability before the task; if missing, stop and prepare the documented isolated environment separately. Do not replace tool execution with model mental arithmetic. No live market feed is needed, but Python is still required.
Worked example: DEMO-CONSUMER
These are fictional nine-month RMB-million figures, not a listed company’s results.
| Check | Previous | Current | Expected observation |
|---|---|---|---|
| Revenue | 1,000 | 1,120 | +12% |
| Net income | 120 | 132 | +10% |
| Operating cash flow | 100 | 92 | -8% |
| Cash flow / net income | 83.33% | 69.70% | Weaker conversion |
Prompt in the AI conversation: “Use cotrading-earnings-review on only these files. First list versions, periods, missing fields and intended output. Then compare profit with cash flow, preserving file and section references. Separate facts, management explanations, inference and unknown.”
The current synthetic report attributes slower collection to payment timing. That is an attributed explanation. Missing aging schedules and subsequent receipts mean it is not a verified full causal account.
Accept or return the result
Inspect the evidence table before the prose. Each original figure needs a location and unit; every calculated percentage needs a formula. Check that receivables at September 30 are not described as nine-month receipts. A missing comparative should yield needs-input, not a guessed growth rate.
A completed limited review can retain unanswered questions. Preserve the input files, report and follow-up list under a fresh run ID. If the host cannot read the files or the model has no quota, record that failure separately; seeing the skill name is not a successful research run.
Limits
This text sample does not guarantee scanned-PDF parsing, comprehensive accounting analysis or any buy/sell decision. Method packages are free; the chosen model and optional tools may cost extra. Files processed by a remote model follow that provider’s settings.
FAQ
Does a selected skill prove the review is done?
No. Check source correspondence and calculations in the actual output.
Can I provide only the latest report?
Yes for a limited reading; comparative changes need comparable prior inputs.