Thulpoff

Skill distillation for AI agents.
Generate, evaluate, and refine SKILL.md files using teacher-student model distillation.

---

Thulpoff captures domain expertise from capable "teacher" models and encodes it into structured skill files that enhance smaller "student" models. It implements the upskill paradigm in Rust — generation, evaluation, and iterative refinement of SKILL.md files compatible with Claude Code, thulp, or any agent supporting the skill format.

Built by DIRMACS.

Install

cargo install thulpoff-cli

# Or as a library
[dependencies]
thulpoff-core = "0.1"
thulpoff-engine = "0.1"
thulpoff-provider = "0.1"

Why Thulpoff?

Small models can match large models on specific tasks — if given the right instructions. But writing those instructions is manual and brittle. Thulpoff automates the loop:

1. Teacher demonstrates — a capable model solves a task
2. Skill extraction — solution patterns are distilled into SKILL.md
3. Test generation — test cases validate the skill works
4. Student evaluation — smaller models are scored with the skill
5. Refinement — failures feed back to improve the skill

No manual prompt engineering. The teacher teaches, the student learns, the skill improves.

Workspace (4 crates)

Crate	What	Tests
thulpoff-core	Types, traits, LlmProvider, CompletionRequest/Response	8
thulpoff-provider	AnthropicProvider (Claude API), NimProvider (NVIDIA NIM)	8
thulpoff-engine	GenerationEngine, EvaluationEngine, RefinementEngine	14
thulpoff-cli	generate, eval, refine, list commands	6

Quick Start

Generate a skill

thulpoff generate \
  "Write an optimized sorting algorithm" \
  --model claude-opus-4-6 \
  --provider anthropic \
  --output ./skills/

Evaluate with a student model

thulpoff eval \
  ./skills/optimized-sorting/SKILL.md \
  --model mistralai/mistral-small-24b-instruct-2501 \
  --provider nim

Refine based on failures

thulpoff refine \
  ./skills/optimized-sorting/SKILL.md \
  --model claude-opus-4-6 \
  --provider anthropic

List available skills

thulpoff list --dir ./skills/

Providers

Provider	Models	Auth
anthropic	Claude Opus, Sonnet, Haiku	ANTHROPIC_API_KEY
nim	Mistral, Llama, Nemotron via NVIDIA NIM	NVIDIA_API_KEY

Architecture

thulpoff/
  crates/
    thulpoff-core/      # types, traits, LlmProvider interface
    thulpoff-provider/  # Anthropic + NIM provider implementations
    thulpoff-engine/    # generation, evaluation, refinement engines
    thulpoff-cli/       # clap CLI with provider selection
  docs/                 # architecture, CLI reference, types, roadmap
  reference/            # upskill Python reference (temporary)

The Distillation Loop

Development

cargo build --workspace
cargo test --workspace
cargo clippy --workspace -- -D warnings

Ecosystem

Project	What
thulp	Execution context engineering — tool discovery, validation, workflows
ares	Agentic retrieval-enhanced server
pawan	CLI coding agent
eruka	Context intelligence engine

Inspiration

• HuggingFace upskill — the original Python implementation
• Claude Code Skills — the SKILL.md format
• We Got Claude to Build CUDA Kernels — the approach demonstrated

License

MIT OR Apache-2.0