Embed badge
Show
Style
[](https://versuz.dev/claude-md/elizaos-eliza)A CLAUDE.md is just a markdown file at the root of your repo. Copy the content below into your own project's CLAUDE.md to give your agent the same context.
npx versuz@latest install elizaos-eliza --kind=claude-mdcurl -o CLAUDE.md https://raw.githubusercontent.com/elizaOS/eliza/HEAD/CLAUDE.md# CLAUDE.md — packages/training Read [`AGENTS.md`](AGENTS.md) first. It is the canonical contract for this directory and binds every agent (Claude or otherwise) working on the Eliza-1 training, quantization, eval, and HF-publishing pipeline. Claude-specific notes: - The repo-wide `CLAUDE.md` at the workspace root applies on top of this file. Read both. The "review-first file writes" and "scope discipline" rules there apply to scripts under this directory too. - When in doubt about runtime mandates (mandatory optimizations, bundle layout, manifest schema, three-mode product shape), consult [`packages/inference/AGENTS.md`](../inference/AGENTS.md). Do not duplicate or paraphrase those rules into training scripts — read them, apply them. - When asked to "skip the eval gate" or "publish anyway", push back. AGENTS.md §6 is explicit: green eval, green kernels, then publish. - The privacy filter is mandatory on every trajectory write path. The repo-wide CLAUDE.md says so; this file confirms it. - New transform scripts (`transform_*.py`, `synthesize_*.py`) MUST run through `validate_corpus.py` before being included in a training run. No raw output → fine-tune in one step.