Synaptic-Tuner: Skill for Claude Code

Case Studies: Implementing the Training Pipeline

Four end-to-end worked examples showing how to take a capability from concept to trained model.

Why Case Studies?

The other skills teach you how to use individual tools:

• synthetic-data-generation — how to run SynthChat
• fine-tuning — how to run trainers
• evaluation — how to run evals
• upload-deployment — how to ship models

This skill shows you how they all connect — the decisions, the iteration, and the order of operations that turn an idea into a trained capability.

The Four Case Studies

Case Study	What It Teaches	Reference
Tool Calling	Structured output training — teaching a model to call APIs with correct syntax, context objects, and parameters	reference/tool-calling-pipeline.md
Essay Style	Creative output training — teaching a model to transform messy brainstorms into structured outlines with voice and personality	reference/essay-style-pipeline.md
Agentic Search	RAG agent training — teaching a model to search a corpus, select relevant documents, and answer grounded in sources	reference/agentic-search-pipeline.md
Harness Distillation	On-policy agent training — a small student acts, a reviewer passes or minimally corrects each step before it runs, and the result becomes per-turn SFT plus DPO pairs	reference/harness-distillation-pipeline.md

The Universal Pipeline

All four case studies follow the same high-level pipeline, but diverge in dataset design and validation:

┌──────────────────────────────────────────────────────────┐
│  1. DEFINE THE CAPABILITY                                 │
│     What should the model do? What does good look like?   │
│     → Rubrics, schemas, specifications                    │
└────────────────────┬─────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────┐
│  2. CREATE TRAINING DATA                                  │
│     How do we generate enough high-quality examples?      │
│     → SynthChat scenarios, handcrafted seeds, self-play   │
└────────────────────┬─────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────┐
│  3. VALIDATE & IMPROVE                                    │
│     How do we ensure quality before training?             │
│     → Schema validation, rubric scoring, manual review    │
└────────────────────┬─────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────┐
│  4. TRAIN                                                 │
│     SFT first (learn the format), then KTO (learn         │
│     preferences), optionally GRPO (optimize rewards)      │
│     → Trainers with YAML config                           │
└────────────────────┬─────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────┐
│  5. EVALUATE                                              │
│     Does the model do what we trained it to do?           │
│     → Evaluator with YAML scenarios                       │
└────────────────────┬─────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────┐
│  6. ITERATE                                               │
│     What failed? Generate more data targeting weaknesses. │
│     → Failure analysis → targeted generation → retrain    │
└──────────────────────────────────────────────────────────┘