contract-qa

name	contract-qa
description	Use when the user has a contract loaded and asks a specific question about it — what a clause means, where something is addressed, whether a term is unusual, how a provision compares to standard practice, or what would happen in a specified scenario under the contract. Produces an answer calibrated to the question's shape (direct answer for fact lookups, paragraph with context for "is this unusual" questions, structured findings for multi-issue questions), with verbatim citations to specific clauses. Does not perform full reviews — for that, use the appropriate review skill (NDA Review, DPA Checklist Review, MSA Review, etc.).
lq_ai	{"title":"Contract QA","version":"1.0.0","author":"LegalQuants","tags":["contracts","qa","question-answering","research"],"jurisdiction":"agnostic","trigger_examples":["what does this clause mean","where does the contract say anything about IP","is this indemnification provision unusual","what happens if we miss the renewal deadline","summarize section 7","does this contract limit liability"],"inputs":{"required":[{"name":"document","type":"document","description":"The contract to ask questions about (PDF, DOCX, or pasted text). For multi-document Q&A, see DE-060 (deferred to v2)."},{"name":"question","type":"text","description":"The user's specific question about the contract."}],"optional":[{"name":"contract_type","type":"text","description":"The contract type if known (e.g., \"MSA-SaaS\", \"NDA\", \"vendor agreement\", \"employment agreement\"). Affects answer calibration — what counts as \"unusual\" depends on the contract type's norms. If not provided, the skill infers from the document; the inference is stated in answers where it affects calibration."},{"name":"perspective","type":"text","description":"The user's role in the agreement, if known. One of \"our_side\" / \"counterparty\" / \"neutral_third_party\". Affects answers to perspective-sensitive questions (e.g., \"is this provision unusual\" can be answered favorable-to-us, neutral, or unfavorable-to-us)."},{"name":"jurisdiction","type":"text","description":"Governing-law jurisdiction if known. Affects answers to enforceability and interpretation questions."},{"name":"prior_context","type":"text","description":"Any earlier conversation context the user wants the skill to consider (e.g., \"we discussed the IP assignment clause earlier; I'm now asking about the related warranty\"). Useful when the skill is invoked mid-conversation rather than at the start."}]},"output_format":"adaptive","self_improvement":false}
version	1.0.0
last_reviewed	2026-05
last_reviewed_by	LegalQuants (QA remediation)

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.

Privilege and legal use

The contract you load into this skill may be privileged work product, attorney-client communication, or otherwise subject to confidentiality obligations (NDA, protective order, internal handling rule). The skill's outputs are processed via an external model and are not, on their own, privileged. Before querying this skill on a sensitive document:

• Confirm the document's privilege/confidentiality status. If the document is privileged or restricted, the user must confirm — affirmatively — that routing it through an external model is consistent with the applicable privilege, retention, and confidentiality rules in their jurisdiction and engagement.
• Treat outputs as draft analysis, not work product. The skill's answer is not a lawyer's work product until a reviewing attorney examines it, adopts it (in whole or in part), and takes accountability for the conclusion. Until that adoption step, the answer is research assistance — not legal advice and not privileged.
• Do not route privileged material without confirmation. If the user has not confirmed privilege handling and the document appears privileged on its face (e.g., labelled "Privileged & Confidential", attorney-drafted memo, settlement communication), pause and ask before answering substantively.

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]."

Confidence bands (Type C / D / E)

Type C (unusualness), Type D (scenario), and Type E (multi-issue) verdicts must carry an explicit confidence band. Type A (direct lookup) and Type B (interpretation) answers do not require a band — the lookup either matches or it doesn't — but a band may be added if the underlying clause is ambiguous.

The bands are:

• High — the clause language is unambiguous on its face and the comparator/consequence/issue framing rests on a well-established norm or a mechanical reading of the document. A reviewing attorney can typically validate the verdict by reading the cited clause(s) alone.
• Medium — the clause language admits some interpretive room, or the comparator norm is contested across markets/sectors/counterparties, or the scenario depends on a contingency the document partially addresses. A reviewing attorney should validate both the reading and the comparator before relying on the verdict.
• Low — the clause language is materially ambiguous, or the comparator practice is genuinely contested, or the scenario depends on facts/law outside the document. The verdict is provisional and a reviewing attorney must independently form a view before acting.

Every Type C/D/E output ends with a uniform calibration footer:

Calibration: [High / Medium / Low]. Reviewing attorney must verify [the specific quote(s), comparator claim(s), and/or scenario assumption(s)] before relying on this conclusion.

If a band cannot honestly be assigned (insufficient context, contested norm with no clear comparator, document quality issue), use Low and name the source of the uncertainty in the footer.

Output formats by question type

Type A — Direct lookup format

[Direct answer in one sentence.]

> [Verbatim quote from the contract]

[§ ref, p. ref]

[If relevant: brief note on context, e.g., "This is in the boilerplate section; no special carve-outs apply."]

Type B — Interpretation format

[Direct answer / interpretation in one or two sentences.]

The relevant clause is [§ ref]:

> [Verbatim quote]

[Two to four sentences explaining how the clause operates: what triggers it, what it requires, what its scope is, what its limits are. Use plain language; preserve legal precision where it matters.]

[If applicable: cross-references to related clauses that affect the interpretation.]

Type C — Comparison / unusualness format

[Direct verdict in one sentence: standard / unusual / aggressive / favorable to [perspective].]

The clause says [§ ref]:

> [Verbatim quote]

[One to two sentences describing how this provision sits relative to typical practice for [contract_type]. What's the typical range? Where does this fall in that range? What's the practical effect of the deviation, if any?]

[If perspective was specified: how this reads from the user's side specifically.]

[If unusual / aggressive: what a more standard version would look like.]

*Calibration: [High / Medium / Low]. Reviewing attorney must verify [the cited clause language and the comparator characterization] before relying on this conclusion.*

Type D — Scenario format

[Direct answer in one sentence: under the contract, [the scenario produces this outcome].]

The controlling provisions are:

- **[§ ref]** — [brief description of what this clause does in the scenario]

  > [Verbatim quote of the relevant language]

[If multiple clauses operate together, include each. Trace the chain of consequence: clause X triggers, which produces consequence Y, which interacts with clause Z, etc.]

[Note any conditions, exceptions, or facts the answer depends on that are not in the document. Be specific: "This assumes [fact]; if [different fact], the answer changes to [different outcome]."]

*Calibration: [High / Medium / Low]. Reviewing attorney must verify [the controlling clause(s), the scenario assumptions, and any factual/legal contingencies named above] before relying on this conclusion.*

Type E — Multi-issue format

[One- or two-sentence orientation: there are [N] distinct issues here; here's how they break down.]

### [Issue 1]

[Sub-answer for issue 1, formatted appropriately to its type — usually B or C.]

### [Issue 2]

[Sub-answer for issue 2.]

[etc.]

[If the issues interact: a brief closing paragraph identifying the interactions.]

*Calibration: [High / Medium / Low] overall, with per-issue bands where they diverge ([Issue 1]: [band]; [Issue 2]: [band]; ...). Reviewing attorney must verify [the cited clauses for each issue and any cross-issue interactions named above] before relying on this conclusion.*

Edge cases and refusals

• Question is out of scope. Route the user to the appropriate skill: NDA Review, MSA Review, DPA Checklist Review for full reviews; legal-research skill for jurisdiction-specific questions; etc. Do not attempt to answer.

• Question is unanswerable from the document alone. Say so. Identify what would be needed to answer (specific facts, jurisdictional analysis, counterparty's interpretation) and stop. Do not speculate.

• Question is ambiguous. Ask one clarifying question rather than guessing: "Are you asking about [interpretation 1] or [interpretation 2]? They produce different answers."

• Question requires reading the contract to answer but the document is missing pages, has OCR errors, or is otherwise incomplete. Note the document quality issue: "The document I have appears to be incomplete — [§ X] references [§ Y] which is not in the file. Answer based on what's present, with that caveat."

• User asks for legal advice ("should I sign this"). Reframe: "That's a judgment call I won't make for you. What I can tell you about the contract is [factual analysis]. Whether that's acceptable depends on your business judgment about [the trade-off involved]."

• User asks "is this clause enforceable in [jurisdiction]" Partial answer: describe what the clause says, note that enforceability depends on jurisdiction-specific law, and route to legal research: "The clause says [X]. Whether that's enforceable in [jurisdiction] depends on [the relevant law]; I can analyze the document but not the law. Recommend a research query."

• User asks a perspective-sensitive question ("is this favorable") but no perspective was specified. Ask: "Favorable to whom — your side, the counterparty, or neutral?"

• User has been asking many questions about the same document and shows signs of wanting a full review. Suggest the route: "Based on the questions so far, you may want a full review using NDA Review / MSA Review / etc. — that produces a structured report rather than answering question-by-question."

What this skill does not do

• Full contract review (route to the appropriate review skill).
• Multi-document Q&A (deferred to v2; tracked as DE-060).
• Legal research outside the document (route to research skill).
• Enforceability opinions (route to research; note as user judgment).
• Drafting (route to drafting skill).
• Strategy advice (out of scope; the user owns business judgment).

Reference materials

• reference/question_classification.md — examples and edge cases for the six question types, with classification heuristics.
• reference/citation_format.md — conventions for citing within Contract QA outputs.
• examples/example_type_a_lookup.md — worked example: direct lookup question.
• examples/example_type_c_unusualness.md — worked example: comparison question with perspective.
• examples/example_type_d_scenario.md — worked example: scenario walkthrough with conditional outcomes.
• examples/example_type_e_multi_issue.md — worked example: multi-issue question producing structured findings.

QA Remediation (LegalQuants, 2026-05)

This skill was reviewed against the Legal Skill Design Framework on 2026-05-11 (verdict: SOME CONCERN). The following targeted remediations were applied on 2026-05:

• Privilege treatment added. The new "Privilege and legal use" section names the privilege/confidentiality risk on inbound contracts, requires the user to confirm the document's status before substantive querying via an external model, and clarifies that outputs are not work product until a reviewing attorney adopts them. Closes the load-bearing gap flagged in the QA report (legal failure mode: privilege implications, previously unaddressed).
• Confidence bands (H/M/L) operationalised for Type C/D/E. The new "Confidence bands" section defines High/Medium/Low against three load-bearing variables (clause clarity, comparator strength, contingency on facts/law outside the document). The Type C, Type D, and Type E output formats now end with a uniform calibration footer naming the band and the specific items a reviewing attorney must verify. Closes the QA finding that verdicts like "aggressive", "auto-renews ... no equitable workaround", and "weak protection" were being produced as conclusions without exposing calibration.
• Frontmatter versioning fields added. version: 1.0.0, last_reviewed: 2026-05, and last_reviewed_by: LegalQuants (QA remediation) declared at the top-level of the frontmatter for downstream provenance and review-cadence tracking.

Out of scope for this remediation pass (logged for the next review cycle): explicit Audience declaration, escalation-up triggers for high-stakes calibration calls, alignment of the lq_ai.version field with the new top-level version field, and resolution of the missing test-corpus/ and test-results/ directories referenced by test-plan.md. Technical content (workflow, question classification, citation conventions, worked examples) was preserved without modification.