🔥 d-skill-forge

Distill procedural skills from frontier model traces into reusable SKILL.md artifacts

---

What is d-skill-forge?

d-skill-forge captures how strong models (Claude Opus, GPT-5) solve tasks, reflects on the execution traces, and produces a structured SKILL.md file. When loaded into a weaker model's context, the skill improves its performance on the same domain — no fine-tuning required.

Task corpus  →  STRONG MODEL  →  Traces  →  REFLECTIVE DISTILL  →  SKILL.md  →  WEAK MODEL  →  +40% score

---

Install

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/d-init-d/d-skill-forge/main/install.sh | bash

Windows (PowerShell)

irm https://raw.githubusercontent.com/d-init-d/d-skill-forge/main/install.ps1 | iex

From source (with pip)

pip install d-skill-forge[tui]

---

Usage

dskillforge

That's it. The TUI launches automatically.

TUI Features

Key	Action
c	Connect a provider
m	Select model
1	Run corpus
2	Extract skill
3	Evaluate
4	Lint
Tab	Next step
q	Quit

CLI Mode

# Run without TUI
dskillforge run --provider groq --model llama-3.3-70b-versatile --corpus tasks.yaml
dskillforge extract --run runs/<id> --provider groq --out skills/SKILL.md
dskillforge eval --skill skills/SKILL.md --corpus tasks.yaml --provider mock
dskillforge lint skills/SKILL.md

---

Providers (15+)

Provider	Type	Auth
Anthropic	Built-in	ANTHROPIC_API_KEY
OpenAI	Built-in	OPENAI_API_KEY
Groq	Preset	GROQ_API_KEY
DeepSeek	Preset	DEEPSEEK_API_KEY
Together AI	Preset	TOGETHER_API_KEY
Fireworks AI	Preset	FIREWORKS_API_KEY
OpenRouter	Preset	OPENROUTER_API_KEY
xAI (Grok)	Preset	XAI_API_KEY
NVIDIA	Preset	NVIDIA_API_KEY
Cerebras	Preset	CEREBRAS_API_KEY
Amazon Bedrock	Built-in	AWS credentials
Google Gemini	Built-in	GOOGLE_API_KEY
Ollama	Local	(none)
LM Studio	Local	(none)
Any OpenAI-compatible	Custom	configurable

Connect in TUI: press c → select provider → paste key → done.

---

How It Works

# 1. Init a project
dskillforge init demo && cd demo

# 2. Run tasks against a strong model
dskillforge run --provider anthropic --model claude-opus-4 --corpus tasks.yaml

# 3. Extract a skill from the traces
dskillforge extract --run runs/<id> --provider anthropic --out skills/SKILL.md

# 4. Evaluate: does the skill help a weaker model?
dskillforge eval --skill skills/SKILL.md --corpus tasks.yaml --provider anthropic --weak-model claude-haiku-4

# 5. Validate
dskillforge lint skills/SKILL.md

Output: SKILL.md

---
name: python-debug
description: Fix common Python runtime errors
version: 1.0.0
source_model: claude-opus-4
triggers: [TypeError, AttributeError, KeyError]
domains: [python, debugging]
---

## Strategy
When encountering a Python error...

## Examples
...

---

Quick Demo (no API key needed)

dskillforge init demo && cd demo
dskillforge run --provider mock --corpus tasks.yaml
dskillforge extract --run runs/<id> --provider mock --out skills/demo/SKILL.md
dskillforge eval --skill skills/demo/SKILL.md --corpus tasks.yaml --provider mock

---

Documentation

• TUI Guide
• Providers
• Authentication
• Concepts
• Architecture
• CLI Reference

---

Contributing

See CONTRIBUTING.md.

License

Apache-2.0