runcardlisted
Install: claude install-skill ruijiezh67/ruijie_skills
# runcard
An agent that is about to spend two GPU-hours has resolved dozens of settings, and the user has seen maybe three of them. The rest come from config defaults, from an earlier run, or from the agent's own guess. That is where wrong checkpoints and overwritten output directories come from, and a prose plan does not expose them: it reads like intent, not like the values that will actually be used.
This skill renders those values, says where each one came from, and asks once.
## 1. Resolve what will actually be used
**Resolve by reading, never by running the thing you are about to confirm.** Executing the entry point to "see what it resolves to" is the launch you were asked to confirm: by the time the card appears, the money is spent. Read the files instead, and use a resolver only when it is explicitly a no-op (`--print-config`, `--cfg job --resolve`, `--dryrun`, `--test-only`); check `--help` first if you are unsure whether a flag trains.
Work out the settings the pending command will really see, in this order of authority: explicit command-line arguments, then the config files that command loads (follow includes and inheritance), then environment variables, then framework defaults. If the project ships a resolver, use it instead of guessing: `python train.py --cfg job --resolve` for Hydra, `accelerate env`, `python -c "import json,yaml;..."` for a plain YAML, or the project's own `--print-config` flag. Resolve interpolations: show `llama3-8b-s1337`, never `${mode