itmegirish/.claude/skills/section-drafter-prompt

SKILL: draft-prompt

Purpose

Build and refine the draft prompt that produces a complete court-ready legal document in a single LLM call. v5.1 uses free-text drafting — the LLM outputs the entire document (not section-keyed JSON, not gap-fill). Exemplar-guided, LKB-informed.

When to Use

• Building or modifying prompts/draft_prompt.py
• Debugging why draft quality is low
• Tuning exemplars or context injection
• Adapting prompt for a new cause type
• Optimizing prompt token count

Architecture Context (v5.1 — what's running)

Single LLM call produces the complete document. No templates, no section keys, no gap-fill markers.

intake_classify -> rag -> enrichment -> draft_freetext_node -> 4 gates -> review

Model: glm-5:cloud with reasoning=ON, temperature=0.7, ~30-50s

---

System Prompt Structure

build_draft_freetext_system_prompt(doc_type, cause_type) builds:

1. Role — "senior Indian litigation lawyer, 25 years practice"
2. Exemplar — loaded via load_exemplar(doc_type) (~1,500 tokens)
3. LKB quality rules — Q1-Q10, COA rules, relief rules (from LKB entry)
4. Substantive rules — R1-R16 (anti-hallucination, citation, formatting)
5. Output instruction — "Write the COMPLETE document as it would appear when filed in court"

Key instruction: plain text output, not JSON, not markdown code blocks.

---

User Prompt Structure

build_draft_freetext_user_prompt(...) includes (in order):

1. User request (facts FIRST — lost-in-middle principle)
2. Extracted facts summary from intake
3. Parties context
4. Evidence context
5. Limitation context — from _build_limitation_context()
6. Verified provisions — from _build_verified_provisions_context()
7. LKB brief — from _build_lkb_brief_context() (acts, limitation, terminology)
8. RAG context — from _build_rag_context() (top chunks, deduped)

---

Context Builders (in draft_single_call.py)

| Builder | What it provides | |---------|-----------------| | _build_limitation_context() | Limitation article, period, description. Handles article == "NONE" | | _build_verified_provisions_context() | List of verified statutory provisions from enrichment | | _build_rag_context() | Top RAG chunks from Qdrant, deduped, scored | | _build_lkb_brief_context() | Primary acts, limitation summary, terminology, anti-hallucination instruction | | _build_procedural_requirements_context() | Court rules, procedural requirements from LKB |

---

LKB Brief Builder

_build_lkb_brief_context() feeds verified legal knowledge to the draft LLM:

• Primary acts with sections
• Alternative acts
• Limitation article + period (or "NO LIMITATION APPLIES" when article=NONE)
• Terminology mapping (conditional, resolved at runtime)
• Anti-hallucination instruction: "Use RAG text, not training memory"

---

Exemplars

Structural exemplars per cause type in exemplars/ directory (~1,500 tokens each):

• Show document structure, section order, heading style
• Use generic placeholders, not specific facts
• Guide LLM on Indian legal drafting conventions

To add a new exemplar:

1. Create file in exemplars/ matching cause type name
2. load_exemplar(doc_type) auto-discovers by naming convention
3. Exemplar included in system prompt

---

Draft Node Flow (draft_single_call.py)

draft_freetext_node(state):

1. Build system prompt (exemplar + LKB + rules)
2. Build user prompt (facts + context)
3. Single LLM call -> raw text output
4. Strip markdown fences if present (defensive)
5. Clean encoding artifacts
6. Append advocate block if absent
7. Collect placeholders
8. Wrap in DraftArtifact -> draft.draft_artifacts[0]
9. Route to evidence_anchoring gate

---

Key Files

| File | What | |------|------| | prompts/draft_prompt.py | System/user prompt builders + exemplar loading | | nodes/draft_single_call.py | draft_freetext_node + all context builders | | exemplars/ | Structural exemplars per cause type | | lkb/civil.py | LKB entries feeding the brief |

---

Rules

• User facts go FIRST in user prompt (lost-in-middle: LLMs attend best to start/end)
• Only cite VERIFIED PROVISIONS — enrichment provides the list
• Exemplar guides structure, LKB guides substance
• Plain text output — no JSON, no markdown code blocks
• Max 5 system instruction rules (Few-Shot > Exhaustive Rules)
• Strip markdown fences defensively (LLM sometimes wraps)
• Handle limitation.article == "NONE" explicitly

Anti-Patterns

• Do NOT use JSON section-key output format
• Do NOT add must_include checklists in prompt
• Do NOT add per-scenario instructions (if partition do X, if money do Y)
• Do NOT dump all RAG chunks raw — use context builders to filter/score
• Do NOT constrain legal reasoning ("must cite Section 65")
• Do NOT hardcode legal content in Python — LKB + RAG + exemplar handles it
• Do NOT exceed prompt budget unnecessarily — compression matters for speed