Legal Drafting Workflow

SKILL: legal-drafting-workflow

Purpose

Execution guide for the drafting pipeline. v5.1 is running. v11.0 is target.

Architecture Reference

• CLAUDE.md: .claude/CLAUDE.md (v11.0 target architecture)
• Skills: /v9-architecture, /schema-builder, /lkb-enrichment, /prompt-builder

How to Run

Run single draft

agent_steer/Scripts/python.exe research/run_draft_live.py

Unit tests

agent_steer/Scripts/python.exe -m pytest tests/drafting/ -v

Benchmark (10 scenarios)

agent_steer/Scripts/python.exe research/run_draft_benchmark.py --draft
agent_steer/Scripts/python.exe research/run_draft_benchmark.py --compare

Pipeline Flow (v5.1 — what's running)

STAGE 1: CONTEXT GATHERING (~12-15s, 1 LLM call)
  intake_classify (qwen3.5:cloud)
    -> domain_router -> enrichment (LKB lookup, 0 LLM, 0 API)

STAGE 2: FREE-TEXT DRAFTING (~30-50s, 1 LLM call)
  draft_freetext (qwen3.5:cloud, reasoning=True)
    LKB-guided, outputs complete court-ready document

STAGE 3: DETERMINISTIC VALIDATION (~0.1s, zero LLM)
  evidence_anchoring -> lkb_compliance -> postprocess -> citation_validator

STAGE 4: REVIEW (~40-70s, 1 LLM call — slim payload, optional)
  review (OpenAI gpt-5.2)
    Slim context: draft text + gate errors + user request (~7K tokens)

Pipeline Flow (v11.0 — target)

STAGE 1: CONTEXT GATHERING (~12-15s, 1 LLM call)
  intake_classify (cause_type + document_type)
    -> domain_router -> enrichment (LKB + document schema)

STAGE 2: DRAFTING (~30-50s, 1 LLM call)
  draft (structured prompt: schema + LKB Layer 1 + Layer 2 + facts)
    LLM drafts complete document from structured context

STAGE 3: DETERMINISTIC VALIDATION (~0.1s, zero LLM)
  evidence_anchoring -> lkb_compliance -> postprocess -> citation_validator

STAGE 4: REVIEW (optional)
  review (OpenAI gpt-5.2)

Key Files

To change draft quality (v11.0 approach)

1. lkb/causes/*.py — LKB entries Layer 1 + Layer 2 (legal data + document components)
2. schemas/*.py — document schemas (section order + instructions)
3. prompts/draft_prompt.py — structured prompt builder
4. nodes/draft_single_call.py — _build_lkb_brief_context() (to be replaced by prompt builder)

To add a new cause type

1. lkb/causes/{file}.py — add entry with Layer 1 + Layer 2 fields
2. lkb/__init__.py — add aliases if needed
3. Done. Works with ALL 12 document types automatically.

To add a new document type

1. schemas/{file}.py — add schema with section order + instructions
2. Done. Works with ALL 92 cause types automatically.

To change models

1. app/config/settings.py — OLLAMA_DRAFT_MODEL, OLLAMA_REVIEW_MODEL
2. Review: now OpenAI gpt-5.2 (changed 2026-03-14)

To change validation gates

1. nodes/evidence_anchoring.py — fact → intake tracing
2. nodes/lkb_compliance.py — act citation + superseded law
3. nodes/postprocess.py — formatting
4. nodes/citation_validator.py — provision verification
5. nodes/civil_decision.py — family consistency gates (7 families)

Debug Checklist

When a draft scores below target:

1. Check intake — did it classify cause_type correctly? Check [INTAKE+CLASSIFY] log
2. Check LKB — did lookup succeed? Watch for [LKB] miss. Check aliases
3. Check enrichment — did limitation resolve? Check [ENRICHMENT] log
4. Check prompt — is LKB data reaching draft LLM? Check _build_lkb_brief_context output
5. Check gates — false positives? Check gate logs
6. Check review — did review fix or break? Check [REVIEW] blocking_issues

Common Issues

v11.0 Migration Status