Start here
Three rules and a minimal setup for using language models in research.
Three rules
- Check every output you use. Each workflow has a How to check the result section. Do not skip it.
- Know what you are sharing. Unpublished data, proposals under review and other people’s manuscripts may not be yours to upload. See Policies.
- Keep a record. Save prompts, model names and dates with your analysis, as you would for any instrument setting.
Minimal setup
- A frontier chat model for reading, writing and reasoning tasks.
- A coding agent for anything that touches code or data: it runs and tests what it writes.
- Your usual tools: ADS, arXiv, Python and Astropy. The workflows plug models into these rather than replacing them.
Reading a workflow
Every workflow starts with a header. The line that matters most is QFLAG:
| Level | Meaning |
|---|---|
| L0 | Draft, not yet verified |
| L1 | Tested by the author on a real case |
| L2 | Reproduced by a second researcher |
| L3 | Checked across models or dates |