Image: GitHub
Why it mattersA developer can now train a small model on how they work with an agent. The dataset is their own session log.
Every session with a coding agent is a worked example. It shows your repositories, your conventions and your tools. A new open skill treats those sessions as training data.
drost-ai/tinker-ai-skill was published on 7 November 2026. It reached 245 stars on GitHub by the next day, with 34 forks. It is an agent skill, under MIT. It turns Claude Code transcripts into a fine-tuned model you run on your own machine. The skill works in Claude Code, OpenAI Codex and OpenCode. Any agent that reads the Agent Skills format can use it.
What the skill does
The README lists a nine-step pipeline. The skill collects transcripts from your Claude projects folder, with checksums. It rewrites them as OpenAI-style messages. It slices them to the model's context length. It splits them into training and holdout sets. It trains a LoRA adapter on Tinker. It samples checkpoints. It merges the adapter into the base model. The last step converts the result to a GGUF file. The file runs in llama.cpp.
Three base models are listed. They are Qwen3.6-35B-A3B, Inkling 2 Small, and Qwen3.8-27B. The README names Qwen3.8-27B as the base of Tess-4.5. The skill calls Tinker's API. Any model Tinker supports will work in principle.
Who pays for what
Tinker is a product of Thinking Machines Lab. Its own page says it charges per token for training, sampling and prefill. It also charges 0.10 dollars per gigabyte per month for checkpoint storage.
The skill's README gives its own figures for one epoch of 64K rows. Qwen3.6-35B-A3B costs about 6 to 55 dollars. The range covers 100 to 1,000 sampled rows. Qwen3.8-27B costs 21 to 185 dollars for the same work. These are the README's own figures. The exact bill depends on the row count and the context length a user picks.
The dataset work is free. It runs locally. The paid part is the training on Tinker's infrastructure. Checkpoint storage is paid too.
Why fine-tune on your own sessions
A general coding model is trained on the whole internet. It does not know one team's conventions. It does not know which linter the team uses. It does not know how the person at the keyboard debugs. A personal model is trained on sessions that contain all three. The skill names the full flow in its own README: a fine-tuning dataset, a LoRA trained on Tinker, and a merged model.
A developer who already keeps months of Claude Code transcripts has the dataset already. The barrier to a first fine-tune is the Tinker bill. The engineering work is in the skill. For a team, the pipeline is a reference. It shows what one such dataset looks like in practice. The merged model runs locally in llama.cpp. Inference is local once the training is paid for.
Source
drost-ai/tinker-ai-skill on GitHub. Tinker by Thinking Machines Lab.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.