contract-qa

Files (14)

• SKILL.md
• comparison-liability.md · acceptance/fixtures/
• lookup-termination.md · acceptance/fixtures/
• scenario-late-payment.md · acceptance/fixtures/
• comparison-liability.golden.json · acceptance/snapshots/
• lookup-termination.golden.json · acceptance/snapshots/
• scenario-late-payment.golden.json · acceptance/snapshots/
• example_type_a_lookup.md · examples/
• example_type_c_unusualness.md · examples/
• example_type_d_scenario.md · examples/
• example_type_e_multi_issue.md · examples/
• citation_format.md · reference/
• question_classification.md · reference/
• test-plan.md

SKILL.md (first lines)

# Contract QA

Answer the user's specific question about a loaded contract. The answer is precise, cited to the document, and shaped by the question — not by a fixed template.

This skill handles question-driven inquiry against a single contract. It is the right skill when the user has a *specific question* and wants a *specific answer*. It is the wrong skill when the user wants a comprehensive review of a contract — for that, route to the appropriate review skill (NDA Review, DPA Checklist Review, MSA-SaaS Review, etc.).

## When this skill applies

Apply when:

- The user has loaded a contract and asks a specific question about its content.
- The user asks "where" something is addressed in the contract.
- The user asks whether a clause is unusual, standard, or aggressive.
- The user asks what would happen under the contract in a specified scenario.
- The user asks for a summary of a specific section, clause, or topic within the contract.

Do not apply when:

- The user asks for a full review or risk assessment of the contract — route to the appropriate review skill.
- The user asks a question that requires legal research outside the document (e.g., "is this enforceable in California" — that needs jurisdiction-specific research, not just document Q&A).
- The user asks a question whose answer depends on facts not in the contract (e.g., "did we breach this" — that depends on facts the skill does not have).
- The user asks for legal opinion (e.g., "should I sign this" — defer to the user's judgment; the skill provides analysis, not advice).
- The user asks across multiple contracts ("compare this MSA with our standard form" or "find any of our SaaS agreements with unlimited liability"). This is multi-document Q&A; deferred to v2.

When refusing to apply, route the user explicitly: "this is better suited to NDA Review" / "this needs research outside the document" / etc.

## Inputs

The skill requires the document and the question. If the user has a contract loaded but no specific question, ask:

> "What would you like to know about this contract?"

Optional inputs (`contract_type`, `perspective`, `jurisdiction`, `prior_context`) calibrate the answer. The skill should:

- **Infer `contract_type` from the document** if not provided, and state the inference in answers where it affects calibration ("Treating this as a SaaS MSA based on the structure...").
- **Use `perspective` when answering perspective-sensitive questions.** "Is this unusual" reads differently from each side. If perspective is not provided and the question is perspective-sensitive, either ask for it or answer from a neutral lens with a note.
- **Reference `jurisdiction` only when the question depends on it.** Most contract Q&A is jurisdiction-agnostic (what does the document say); jurisdiction matters for interpretation and enforceability questions.

## Workflow

The workflow has three steps. Each step's behavior depends on the question type identified in step 1.

### Step 1: Classify the question

Read the user's question and classify it into one of these types. The classification drives the output format (adaptive output is the whole reason this step matters).

**Type A: Direct lookup.** The user wants to know what the contract says about a specific topic.
- "Where does the contract address indemnification?"
- "What's the term of this agreement?"
- "What does Section 7 say?"
- "Does this contract have an arbitration clause?"
- Output format: **direct answer with verbatim quote and citation.** Brief.

**Type B: Interpretation.** The user wants to understand what a clause means, how it operates, or how to apply it.
- "What does this indemnification clause actually require?"
- "How does the renewal mechanism work?"
- "If we terminate, what survives?"
- Output format: **explanation paragraph with verbatim quote and citation.** Plain language; preserve precision.

**Type C: Comparison / unusualness.** The user wants to know whether a provision is standard, unusual, aggressive, or favorable.
- "Is this limitation of liability unusual?"
- "How does this non-compete compare to standard?"
- "Is the IP assignment scope aggressive?"
- Output format: **paragraph with context.** Quote the provision; describe how it sits relative to typical practice for the contract type; calibrate to perspective if specified.

**Type D: Scenario.** The user wants to know what would happen in a specified factual scenario.
- "What happens if we miss the renewal notice deadline?"
- "If they breach, what are our remedies?"
- "If we want to assign this contract, what do we have to do?"
- Output format: **scenario walk-through.** Cite the controlling clauses; trace the consequence; note where the answer depends on facts the skill does not have.

**Type E: Multi-issue.** The user asks about multiple topics in one question, or asks about a topic that has multiple distinct aspects.
- "What are the IP and confidentiality terms?"
- "Walk me through the termination provisions."
- "What protections do we have if they go bankrupt?"
- Output format: **structured findings.** One subsection per issue, each with quote and citation.

**Type F: Out of scope.** The question is one the skill should not answer (full review, legal research, scenario depending on external facts, opinion).
- Route to the appropriate path; do not answer.

If the question does not cleanly fit one type, default to the type that produces the most useful answer (typically B or E).

### Step 2: Find the relevant content in the document

For every question type except F:

- Locate the clauses, sections, or content in the document that answer the question.
- For Type A and B, this is typically one or a few specific clauses.
- For Type C, this includes the clause being asked about plus any related provisions (e.g., a limitation of liability question may need the LoL clause plus any carve-outs in the indemnification section).
- For Type D, this includes the controlling clause plus any conditions, exceptions, and consequence provisions that the scenario triggers.
- For Type E, this includes all clauses relevant to all issues asked about.

If the relevant content is not in the document, say so explicitly: *"The contract does not address this directly. The closest provisions are [X], which suggests [interpretation], but the question is not resolved by the document."*

If the relevant content is ambiguous, say so: *"The contract is ambiguous on this point. Section X says [quote A], which would suggest [interpretation 1], but Section Y says [quote B], which would suggest [interpretation 2]. The contract does not resolve the conflict."*

### Step 3: Produce the answer

Format the answer per the question type identified in Step 1.

**For all formats:** Lead with the answer. The user asked a question; the first sentence should answer it. Citations support the answer; they do not bury it.

**Verbatim quotes** must be exact. If you cannot quote verbatim, paraphrase explicitly ("The clause says, in substance, ..."). Do not produce reconstructed quotes that look verbatim but aren't.

**Citations** include the section/clause reference and (where the document supports it) the page reference. Format: `[§4.2(b)]` or `[§4.2(b), p. 7]`.

**Length is proportional to question shape.** A Type A question gets a sentence. A Type C question gets a paragraph. A Type E question gets one subsection per issue. Do not pad.

**Always identify what you cannot answer.** If part of the question is out of scope, or depends on facts not in the document, name that explicitly at the end: *"The skill cannot determine [X] from the document alone — that depends on [the user's facts / jurisdiction-specific law / counterparty's behavior]."*

## Output formats by question type

### Type A — Direct lookup format

```markdown
[Direct answer in one sentence.]

> [Verbatim quote from the contract]

From LegalQuants/lq-ai, Apache-2.0.

More from LegalQuants/lq-ai