Bdistill Export

When to use

• Make your AI tool domain-reliable — paste validated rules into Claude Project/Cursor/Copilot
• Feed rules to a deterministic agent — export as Python module with RULES + build_prompt()
• Export JSON for bdistill-operationalize to contrast against live data
• Generate fine-tuning JSONL for LoRA training (alpaca/sharegpt/openai formats)
• Share with non-technical team — Excel with quality color-coding or audit checklist CSV

Input contract

required:
  domain: string          # Knowledge base domain name (e.g. "aml-compliance")
  format: enum            # prompt | harness-json | harness-python | excel | checklist | training-jsonl

optional:
  platform: enum          # claude-project | cursor-rules | copilot-instructions | chatgpt-custom | generic
  training_format: enum   # alpaca | sharegpt | openai
  min_quality: float      # Minimum quality score threshold (default: 0.7)
  max_rules: int          # Maximum number of rules to include (default: 50)
  max_context: int        # Maximum number of context entries to include (default: 20)

Output contract

format: File on disk. Path depends on format:
  prompt:         data/knowledge/exports/prompts/{domain}-{platform}-{date}.md
  harness-json:   data/knowledge/exports/harness/{domain}_{date}.json
  harness-python: data/knowledge/exports/harness/{domain}_{date}.py
  excel:          data/knowledge/exports/{domain}.xlsx
  checklist:      data/knowledge/exports/{domain}-checklist.csv
  training-jsonl: data/knowledge/exports/training/{domain}-{date}-train-{fmt}.jsonl  # + val split

returns:
  path: string            # Absolute path to exported file
  format: string          # Format used
  entries_exported: int   # Total entries written
  stats:
    rules_count: int
    context_count: int
    quality_distribution: object   # e.g. {A: 12, B: 8, C: 3}

Format comparison table

With bdistill MCP (full power)

• prompt: Call bdistill_export_prompt with domain, platform, min_quality, max_rules, max_context
• harness: Call bdistill_export_harness with domain, format (json/python-dict/python-dataclass), min_quality, max_rules
• excel: Call bdistill_export_excel with domain, min_quality
• checklist: Call bdistill_export_checklist with domain
• training: Call bdistill_training_export with domain, format (alpaca/sharegpt/openai)

Standalone (no dependencies)

1. Read entries from data/knowledge/base/{domain}.jsonl + data/rules/base/{domain}.jsonl
2. Filter by min_quality, sort by confidence descending
3. Partition into rules (entries containing IF/THEN/WHEN/THRESHOLD) and context (everything else)
4. Cap at max_rules and max_context
5. Format according to chosen format:
   • prompt: Generate markdown with ## Rules section (IF-THEN formatted) + ## Context section (Q&A formatted). Add platform-specific header.
   • harness-json: Write JSON with metadata, rules array, context array
   • training-jsonl: Convert to instruction/output pairs, split 80/20 train/val

Platform paste instructions

• Claude Project: paste into Custom Instructions field
• Cursor: save as .cursor/rules/{domain}.mdc
• Copilot: append to .github/copilot-instructions.md
• ChatGPT: paste into Custom GPT Builder Instructions (8K char limit -- export truncates)

Edge cases

• Domain has no entries: return error "No entries found for domain: {domain}"
• All entries below min_quality: lower threshold to 0.5 and warn
• Training export < 50 entries: warn "LoRA fine-tuning works best with 200+ entries"

Example

Export "aml-compliance" as claude-project prompt:

• Input: domain="aml-compliance", format="prompt", platform="claude-project"
• Output: Markdown with 15 rules + 8 context entries, saved to exports/prompts/

Composes with

• bdistill-operationalize: harness-json output is loaded as rules_path
• Any AI tool: prompt output is pasted into the tool's instruction field
• bdistill-extract: builds the KB that this skill exports