Two reusable templates for different stages of the work.
Use 1-write-the-task-prompt.md in a planning chat. Add your rough idea, available information, desired result and limits. GPT's job here is to write a complete brief for the model that will do the work. It should ask about essential gaps rather than invent missing details.
Review the brief before using it. You can give it to Claude or another GPT model. The /goal step below is specifically for Codex.
An independent test of Jev 1.13 against historical NASA archive classifications.
I took 8,054 historical Kepler Objects of Interest from the NASA Exoplanet Archive, hid the archive disposition (classification), and asked Jev to classify each object as CONFIRMED, CANDIDATE or FALSE POSITIVE. Jev is TypeSafe's model for choosing among predefined options; Kepler Objects of Interest are signals found in the Kepler space telescope's observations of dimming stars.
No fine-tuning. No examples from this dataset. One prediction per object, stored before its archive label was read.
8,054 calls · 0 failures · 322 ms median · ~$0.34 estimated
A third-party skill is not a document. It is a set of instructions your agent will follow with your local permissions, plus whatever scripts, hooks and configs ship next to it, plus whatever it tells the agent to install. Treat it as untrusted until you have read what it can make the agent do.
This prompt runs a read-only audit inside Claude Code. It maps what the skill would execute, install, read, send over the network and change in your Claude config, follows the files SKILL.md references, treats the skill's own text as data rather than instructions, and returns a verdict with file:line evidence and the commit it reviewed. It installs and executes nothing.
Replace <SKILL_URL_OR_PATH> and paste the whole block into Claude Code.