メニュー
Use weekly. The teacher log captures every non-local Claude call — curate the best ones into training data so the local model learns from your actual work.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | teacher-trace-curation |
| description | Use weekly. The teacher log captures every non-local Claude call — curate the best ones into training data so the local model learns from your actual work. |
| version | 1.0.0 |
| author | kbot |
| license | MIT |
| metadata | {"kbot":{"tags":["training","fine-tuning","mlx","local-model","dataset"],"related_skills":["autopoiesis-loop","skill-self-authorship"]}} |
Every time kbot calls Claude, the prompt + response is written to ~/.kbot/teacher/traces.jsonl. Left alone, this is just a log. Curated, it's the dataset that teaches the local model to answer your questions without touching the API.
ONLY SUCCESSFUL, CORRECTED, AND USER-APPROVED TRACES ENTER THE DATASET.
Failed traces teach the model to fail. Garbage in is not "more data."
kbot train-self --mode default --max-examples 500 --iters 200 --num-layers 8 — curates + fine-tunes in one pass. The curator runs first, scores traces, and writes ~/.kbot/teacher/dataset-default.jsonl.ollama run kernel-self:<timestamp> on a task from the last week. Compare against the Claude baseline.For longer cycles of evaluation + retraining, use kbot train-cycle which chains curate → train → evaluate → merge across multiple iterations.
Was this answer used without correction? The curator scores partly on: no correction in the next 5 turns, no follow-up question asking for clarification, no user rephrasing. Approved-by-silence is the strongest endorsement.
A local model that answers "how do I deploy this?" using your deploy flow, not Anthropic's generic best practice. Your infrastructure, your naming, your conventions, your past decisions. That's what the local model becomes over weeks.
Training on everything. Larger datasets with noisy data fine-tune worse models than small curated datasets. 200 excellent examples beat 2,000 mediocre ones every time.
KBOT_TEACHER_LOG=1 in ~/.zshrc keeps the logger always-on.launchd plist at ~/Library/LaunchAgents/com.kernel.kbot-train-self.plist runs the curation + training weekly.kernel-self:<timestamp> in Ollama and becomes available to every kbot command via --model kernel-self:latest.Creates distinctive editorial art direction for kernel.chat issues before implementation. Use when choosing an issue's visual concept, typography, composition, cover system, media grammar, responsive transformation, or motion intent; when a page feels generic or lacks character; or when original image and video assets need one coherent visual language.
Audits kernel.chat pages for responsive visual craft, accessibility, motion safety, interaction truth, runtime health, and production rendering. Use when reviewing an issue or artifact, checking whether a design is finished, comparing desktop and mobile, validating reduced motion or print, finding overflow and broken media, or proving that a published route matches the build.
Guides the design, layout, styling, and interactivity of kernel.chat editorial spreads and standalone artifact editions. Use this skill when editing or creating pages, React components, CSS files, or single-file HTML artifacts in the kernel.chat project to enforce the POPEYE-inspired bilingual design grammar and the interaction-language/artifact-language laws.
Use this skill to generate well-branded editorial interfaces and assets for kernel.chat ("Magazine for City Coders" — an independent magazine that is an unnamed homage to POPEYE), either for production or throwaway prototypes/mocks/etc. Contains essential design guidelines, colors, type, fonts, assets, and editorial UI components for prototyping.
Hand off a model from Blender (via BlenderMCP) into Unity (via MCP for Unity) — export the current Blender model, import it through import_model_file, and place it in the open scene. Use when the user has BlenderMCP and MCP for Unity both connected and wants to bring a Blender model into Unity. Does NOT drive Blender's own generators; BlenderMCP owns how the model got into Blender.
Generate images/videos/3D assets/audio via Higgsfield AI. Defaults: GPT Image 2 for image/design/text, Seedance 2.0 for video, Nano Banana 2/Lite/Pro for character/reference images, Marketing Studio for ads, Seed Audio 1.0 for audio. Use when: "generate an image", "make a video", "animate this photo", "image-to-video", "edit/stylize/remix this image", "reframe this video", "edit this video from a sketch", "create a 3D model/GLB", "create a sound effect", "make music", "text-to-audio", "create an ad", "make a UGC video", "unboxing", "presenter video", "import product from URL", or "analyze video virality". Supports image-to-3D (`multi_image_to_3d`), text-to-audio/music (`seed_audio`), workflow generation (`draw_to_video`, `reframe`), Marketing Studio, and Virality Predictor (`brain_activity`). Chain with higgsfield-soul-id for face/identity consistency. NOT for: Soul Character training (use higgsfield-soul-id), product photoshoots, marketplace listing cards, text/chat/TTS tasks.