drost-ai/tinker-ai-skill is a repository that explains how to organize the session conversations of a coding agent into training material and fine-tune an agent model with it. It can be used with Claude Code, Codex, and OpenCode, and it covers the flow of training LoRA on Tinker and exporting the result as a GGUF for local execution. There is no public release information. The repository description states that it is an agent skill for fine-tuning a model with Claude Code sessions.
Instead of extracting only code from the conversation, this procedure treats the context, including tool calls, as training material. The README introduces nine steps connecting transcript collection and anonymization, splitting, training, evaluation, merging, and deployment, along with the related scripts. It takes session-level de-identification and the continuity of conversational context into account, and it also provides a way to dry-run the entire procedure at no cost using synthetic conversations. The documentation says this procedure was used to build Tess-series models.
However, no independent test directory could be confirmed in the repository, so the procedure described in the document alone does not allow a judgment about separate verification results. Also, since it is explained to be designed to stop before the steps that incur cost, the entire flow it describes does not mean that actual training and deployment are carried out at no cost. Therefore, this repository presents a reproducible order of work, but it is closer to procedure guidance whose independent test evidence is unconfirmed.