Your first CoTrading research run in Codex CLI
Start with the text teaching example, not a large folder of private documents. The first objective is to establish a repeatable input-to-output path before asking a model to interpret more material.
Prepare a small run
Follow Codex CLI installation. OpenAI’s skills documentation describes discovery and explicit invocation; it does not certify CoTrading packages. The personal skill folder is separate from the research folder and its reports.
Prepare Python 3.12 for the current 0.1.0 A-batch methods S01/S02/S04/S05, including this text-based S01 task. Their bundled deterministic tools need only the standard library and do not access the network for teaching inputs. The read-only check py -3.12 --version does not install anything or certify a task. If unavailable, stop tool execution and arrange an isolated environment separately; do not change global settings or substitute model mental arithmetic.
| Record | Example | Check |
|---|---|---|
| Method | cotrading-earnings-review, package version | Read the installed manifest |
| Inputs | Two synthetic reports and question | Match names and periods |
| Output | reports/new-run-id | Keep older runs intact |
| Scope | Listed files only | No surprise data replacement |
| Status | completed, needs-input or failed | Must match actual output |
Copy the included example into a folder you choose. Keep its synthetic labels. Open the Codex session there. In the conversation, explicitly mention $cotrading-earnings-review; the dollar-prefixed mention is not a PowerShell command. If it cannot be selected, resolve discovery before proceeding.
A two-stage prompt
First ask: “Identify the selected skill version, input files, output directory, dependencies and missing information. Read only; do not install or fetch anything.” Compare the response with the folder.
Then request the limited review: “Compare the two synthetic reports. Keep source sections and unknowns. Write into a new run directory and do not modify the inputs.” New dependencies or external access should be considered separately rather than hidden inside a research request.
The fictional cash-flow check should retain 100→92 and 120→132, producing -8% operating cash-flow growth and 69.70% current conversion. If prose sounds plausible but the evidence ledger lacks these original values, return the task for correction.
Diagnose the first divergence
If discovery fails, inspect installation level and name. If input loading fails, inspect file readability. If arithmetic differs, inspect units and period before rerunning the model. If quota is exhausted, save inputs and the incomplete status; changing the report wording cannot turn failure into success.
Record host version, method version, input digest and expected-versus-actual checkpoints. This makes a future update testable without publishing private material.
Boundaries
The first text task proves only its observed scope. It does not prove OCR, arbitrary spreadsheets, real-time market access or every host version. Packages are free; a model subscription or provider use may cost extra. Configure credentials in the host’s supported settings, never this website.
FAQ
Do I run the dollar-prefixed name in PowerShell?
No. Mention the skill in the Codex conversation.
Can I reuse the installation directory for reports?
Use a separate work folder so package updates cannot overwrite research.