proposer — Claude Code prompt: purpose and setup
proposer is a Claude Code and ChatGPT prompt that tunes training throughput in InfraRSI by proposing configuration and overlay changes while preserving fixed task settings.
no license ★ 99 updated 2026-11-12
.claude/commands/proposer.mdWhat proposer does
proposer searches for settings that reduce full-step training time in InfraRSI with Megatron-Core 0.18.2. Each proposal is a single JSON action, and the harness measurement decides whether it qualifies; speeding up one component alone does not count.
Changes are limited to configuration fields and overlay files, with a shared cap on edited lines. The model, data order, loss, optimizer math, precision, and other task constants must stay fixed. Overlays also cannot alter timers or collective operations.
The prompt requires reading the previous run’s receipt before proposing another candidate. Failed checks consume GPU-hours, while some of the budget is held for the final scoring pair. It directs the agent to stop when that budget is nearly used or further gains look unlikely.
What it helps with
- reduce full-step training time
- tune Megatron-Core settings within fixed task constraints
- compare a candidate with results from an earlier run
- save lessons from successful and failed configurations
Example prompts for the agent
- Tune full-step training time in InfraRSI without changing the task constants.
- Propose one testable setting based on the previous run’s results.
- Save findings about slow and fast configurations, then name the best valid candidate.
How to install proposer
Manually: put the files in place
.claude/commands/proposer.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about proposer
How do I install proposer in Claude Code?
Place the prompt file at .claude/commands/proposer.md.
What does proposer do?
It suggests configuration and overlay changes to reduce full-step training time in InfraRSI. The harness measurement determines the result.
Which settings must proposer leave unchanged?
The model, data and its order, loss definition, optimizer math, precision, and other task constants are fixed. Overlays cannot modify timers or collective operations.