itmegirish/.claude/skills/template-builder

SKILL: exemplar-builder

Purpose

Create, validate, and maintain document schemas and LKB Layer 2 data for the v11.0 scalable drafting pipeline.

v11.0 approach: No exemplar documents in prompts. Instead: LKB 2-layer data + document schema → structured prompt → LLM drafts.

When to Use

• Creating a new document schema (e.g., written_statement, appeal_memo)
• Enriching LKB entries with Layer 2 data (available_reliefs, jurisdiction_basis)
• Reviewing schema quality against CPC rules
• Adding a new document type to the pipeline
• Debugging why draft structure is wrong

Architecture Context (v11.0)

LKB Layer 1 (legal knowledge) + LKB Layer 2 (document components) + Document Schema
  → Structured prompt (3 sections, ~1,500 tokens)
  → LLM drafts complete document (1 call)
  → Gates validate (0 LLM)

No template engine. No exemplar files. Just better context to the same LLM.

---

Document Schema Directory (v11.0)

app/agents/drafting_agents/schemas/
├── __init__.py          # DOCUMENT_SCHEMAS registry + lookup
├── trial_court.py       # plaint, written_statement, rejoinder, counter_claim
├── applications.py      # interim_application, condonation, set_aside_ex_parte, caveat
├── appellate.py          # appeal_memo, revision_petition, review_petition
└── execution.py          # execution_application

---

Document Schema Format

Each schema defines structure for ONE document type, independent of cause type:

{
    "code": "written_statement",
    "display_name": "Written Statement",
    "filed_by": "defendant",
    "cpc_reference": "Order VIII",
    "annexure_prefix": "D-",
    "verification_type": "defendant",
    "signing_format": "Defendant through Advocate",

    "sections": [
        {"key": "court_heading",          "instruction": "Court name, place, suit number"},
        {"key": "parties",                "instruction": "Defendant and Plaintiff details"},
        {"key": "preliminary_objections", "instruction": "Limitation, jurisdiction, non-joinder objections"},
        {"key": "parawise_reply",         "instruction": "Reply to EVERY plaint paragraph: ADMITTED/DENIED/NOT ADMITTED"},
        {"key": "additional_facts",       "instruction": "New facts supporting defence, not in plaint"},
        {"key": "legal_grounds",          "instruction": "Statutory and legal basis for defence"},
        {"key": "prayer",                 "instruction": "Dismiss suit with costs"},
        {"key": "verification",           "instruction": "Verification on oath"},
    ],

    "filing_rules": {
        "court_fee": False,
        "filing_deadline": "30 days from service (max 120 days — Order VIII Rule 1)",
        "vakalatnama": True,
    },
}

---

LKB Layer 2 Data Format

Each LKB entry gets document component fields alongside existing legal knowledge:

"breach_of_contract": {
    # Layer 1 — Legal knowledge (already exists)
    "primary_acts": [...],
    "limitation": {...},
    "facts_must_cover": [...],

    # Layer 2 — Document components (NEW)
    "available_reliefs": [
        {"type": "damages", "statute": "S.73 ICA", "prayer_text": "decree for damages of Rs.{{AMOUNT}}"},
        {"type": "interest_pre_suit", "statute": "S.34 CPC", "prayer_text": "interest at {{RATE}}% per annum from {{BREACH_DATE}}"},
        {"type": "interest_pendente_lite", "statute": "Order XX Rule 11 CPC", "prayer_text": "pendente lite and future interest"},
        {"type": "costs", "statute": "S.35 CPC", "prayer_text": "costs of the suit"},
    ],
    "jurisdiction_basis": "Section 20 CPC — where cause of action arose",
    "valuation_basis": "Amount of damages claimed",
}

---

Creating a New Document Schema

1. Identify CPC/statute reference for the document type
2. List all required sections in court-standard order
3. Add per-section instruction (what goes in that section)
4. Set filed_by, annexure_prefix, verification_type
5. Add filing_rules (deadline, court fee, etc.)
6. Verify section order against actual court practice
7. Test by running pipeline with the new document type

---

Enriching LKB with Layer 2

1. For each LKB entry, add available_reliefs with prayer_text
2. Add jurisdiction_basis (which CPC section determines jurisdiction)
3. Add valuation_basis (what determines suit valuation/court fee)
4. Verify prayer_text matches actual court prayer format
5. Test by running pipeline — check if prayer output uses exact prayer_text

---

Key Files

| File | What | |------|------| | schemas/ | Document type schemas (NEW) | | lkb/causes/ | LKB entries (enrich with Layer 2) | | lkb/causes/_helpers.py | _entry() schema definition | | prompts/draft_prompt.py | Structured prompt builder | | nodes/draft_single_call.py | _build_lkb_brief_context() (to be replaced) |

---

Rules

• Schema defines STRUCTURE — not legal substance (LKB does that)
• One schema per document type (not per cause type)
• Schema is independent of cause type — same schema works for ALL 92 causes
• available_reliefs.prayer_text must match actual court prayer wording
• Verify schemas against CPC/statute text — wrong structure = court rejects
• Keep sections in court-standard order for that document type

Anti-Patterns

• Do NOT create per-cause-type schemas — schema is per document type only
• Do NOT put legal knowledge in schemas — LKB handles that
• Do NOT use exemplar files in prompts — structured data replaces exemplars
• Do NOT hardcode section order in engine.py — schema drives order
• Do NOT mix Layer 1 and Layer 2 fields — keep them conceptually separate