Bdistill Skills — AI skill for Claude Code

bdistill-skills

**Everyone is building AI agents. But generic agents make generic decisions.**

Your agents need niche domain knowledge to be actually useful — the thresholds, rules, and edge cases that separate a junior from a senior in your field. That knowledge exists inside LLMs, but it's trapped in ephemeral chat sessions. bdistill-skills extracts it into structured, validated knowledge bases that make your agents smart on your niche.

7 composable protocols that distill real intelligence from the AI models you already use — decision rules with numeric thresholds, validated Q&A pairs, structured predictions with evidence. The distilled knowledge persists as a searchable, quality-scored KB that your AI agents consume downstream: specialized assistants, monitoring systems, recommendation engines, fine-tuned local models, or just a Claude Project that finally gives consistent domain answers.

Works with Claude Code, VS Code Copilot, Cursor, Codex CLI, Windsurf, Cline, and 30+ other agents. No API key. No MCP server. No dependencies.

Quickstart (30 seconds)

# 1. Install (pick your tool)
npx skills add FrancyJGLisboa/bdistill-skills          # auto-detects your AI tools
# OR: git clone https://github.com/FrancyJGLisboa/bdistill-skills.git   # manual

# 2. Extract — tell your AI agent:
"Extract rules about [your niche topic]"

# The agent follows the extraction protocol automatically:
# → generates targeted questions → answers them → challenges its own claims
# → scores quality → writes validated rules to data/rules/base/your-domain.jsonl

# 3. Validate — immediately after extracting:
"Validate the KB"

# The agent re-asks every numeric claim 5 different ways + checks structural stability.
# Numbers that vary → flagged as hallucinated. Conditions that appear/disappear → flagged as unstable.
# Only stable entries survive. This is what separates a trusted KB from a chat log.

# 4. Use it:
"Export as system prompt for Claude Proje