tinker-ai-skill Turns Claude Code Sessions Into a Local Model
An MIT-licensed Agent Skill packages the pipeline the author says built Tess-4 and Tess-4.5, from raw transcripts to a quantized GGUF.
A skill that fine-tunes a model on your own coding sessions
The repository drost-ai/tinker-ai-skill had 323 stars and 39 forks on GitHub as of 2026-11-14, seven days after it was created on 2026-11-07. It is an MIT-licensed Agent Skill, written in Python, that lets a coding agent fine-tune model on Claude Code sessions: turn stored transcripts into a supervised fine-tuning dataset, train a LoRA on Tinker, merge it and export a GGUF that runs in llama.cpp locally. The author is Migel Tissera, creator of the Tess models.
The repository says the Tess family was built from the author's own Claude Code sessions with this exact method and released as open weights on Hugging Face, including Tess-4-27B with a GGUF and an EAGLE-3 draft, and Tess-4.5-27B with a GGUF. It calls the published code the cleaned-up version of that pipeline, and names Tinker and tinker-cookbook as products of Thinking Machines Lab, with the skill an independent guide built on them.
What is in the repository
One SKILL.md with a nine-step workflow and its decision points; eleven Python and shell scripts built on tinker-cookbook and llama.cpp; six reference notes on the transcript format, anonymization, chunking, training settings, export and serving; and a synthetic transcript generator for dry runs. Install by cloning into Claude Code's ~/.claude/skills/, Codex's ~/.agents/skills/ or OpenCode's ~/.config/opencode/skills/. OpenCode also reads the other two, and the source says any agent supporting the Agent Skills standard works the same way.
Nine steps, and the failures behind each rule
Collect, clean, slice, compact, split, train, check, merge, ship. Collection copies transcripts verbatim with checksums and skips subagents and tiny sessions. Cleaning replaces every credential by value and gives each session its own random identity, failing rather than writing a file that still leaks. Slicing cuts to the context length measured with the model's real renderer and never splits a tool call from its result. Training keeps tool definitions and the system prompt on every row, and the offline smoke test prints the token count before launch.
The repository lists the failures behind those rules: a single constant fake world was memorized and later asserted over the real environment; rows without tool definitions and a system prompt regressed tool use even though NLL improved; a tokenizer change moved rows by 14% and overflowed the budget; two collapses had monotonically falling loss; and a serving template that strips earlier reasoning, or llama.cpp's default min-p, turned a good checkpoint into garble.
Cost, requirements and the retention trap
Tinker bills per token, and the repository gives ballpark figures at November 2026 prices. One epoch over 100 rows (about 5M tokens) comes to roughly $6 on Qwen3.6-35B-A3B, $8 on Inkling 2 Small and $21 on Qwen3.8-27B; 1,000 rows (about 45M tokens) comes to roughly $55, $70 and $185. Dense models cost more per token than MoE models, and Qwen3.8-27B is the base of Tess-4.5. Dataset steps run locally and are free.
You need Python 3.11+, uv or pip, tinker-cookbook>=0.6.0, wandb and a Tinker API key. Steps 8 and 9 need a built llama.cpp checkout and disk for the base model plus the merged copy. Claude Code deletes transcripts after 30 days by default, so set cleanupPeriodDays to 3650 first.
Limits and open questions
The source contains no independent benchmark of a model trained with the skill; the Tess releases are the author's own. The cost table is an estimate that points to Tinker's price list rather than reproducing it, and how long a real run takes is not stated. Anonymization is a policy, not a guarantee, and the repository says to review anonymized rows before training. Data, runs, exports and .env are git-ignored, and it warns against publishing someone else's transcripts as a dataset.
What it means for building with AI
Fine-tuning on your own sessions is a different lever from prompting, and not the first one to reach for: the dataset work and the evaluation take more care than the training run. For teams building an app rather than training a model, the practical knob is model choice inside a builder, covered in Draftship's documentation on using the AI chat and choosing a model. Draftship is an AI app builder and is not affiliated with this project.
Primary sources
- drost-ai/tinker-ai-skill: Agent Skill for fine-tuning an agentic model on your own Claude Code sessions with Tinker. Works in Claude Code, Codex and OpenCode. · GitHub trending AI repositories, Nov 14, 2026
This article was drafted with AI assistance from the primary sources listed below, then reviewed and approved by a Draftship editor.