A Claude Skill that takes an executed contract and produces an audience-tuned one-page summary — exec, ops, or finance — with explicit “Ambiguous — see legal” fallbacks and a Counsel review required block when the contract trips an escalation trigger. The skill is a reading aid for non-lawyers, not a substitute for the contract or for counsel. It exists to shorten the time between “a contract landed in my inbox” and “I know what I have to do about it” without paving over the nuance the contract turns on.
The artifact bundle lives at /artifacts/contract-summary-skill/: the SKILL.md is the entry point and the three files under references/ are the editable scaffolding the skill loads on every run — the audience question library, the per-audience format templates, and the escalate-to-legal trigger list.
When to use
Drop this skill in when a non-lawyer has to act on a contract:
An exec deciding whether to sign off on a deal whose commercial shape they understand but whose paper they have not read.
A CFO modeling cash flow over the initial term — payment cadence, escalators, take-or-pay, FX risk.
An ops owner who has just inherited responsibility for delivering against a contract someone else negotiated, and needs to know what reports, notices, and SLAs are on the calendar.
A procurement lead twelve months before renewal, building the notice-deadline tracker.
An M&A diligence team that needs a one-page summary of every material contract in the target’s portfolio.
The output is a Markdown summary with §references on every line. Readers can scan the gist in 60 seconds and jump to the cited section of the contract for anything they want to verify.
When NOT to use
Do not use this skill — or any LLM summary — for these:
Legal advice. The skill summarizes; it does not opine on enforceability, defensibility, or whether to litigate. Counsel does that work, against facts the contract does not contain.
Dispute analysis. “Did they breach §7.2?” depends on what happened, what was communicated, what was waived. A summary flattens the nuance the dispute turns on. Read the contract with counsel, not a one-pager.
Privilege determination. Whether a document is privileged is a function of who saw it, in what capacity, and for what purpose. None of that is in the contract.
Non-Tier-A AI vendors. Run this on Claude (with a workspace or team plan whose data-handling posture you have reviewed). Do not pipe contracts — which contain commercial terms, counterparty PII, and sometimes regulated data — through consumer-grade LLMs.
Unsigned drafts treated as executed. Summaries written from redlines are wrong by the time the contract is signed. The skill flags drafts and prepends a stale-summary warning, but the safer default is to wait until you have the executed copy.
Setup
Drop the bundle. Copy /artifacts/contract-summary-skill/ into your Claude Code skills directory at ~/.claude/skills/contract-summary/ (or upload as a Skill in a Claude.ai project). The SKILL.md is the entry point; Claude loads the three files under references/ automatically when the skill runs.
Edit the reference files for your firm. The defaults in references/1-audience-question-library.md, references/2-summary-format-by-audience.md, and references/3-escalate-to-legal-triggers.md are starting points — replace them with your firm’s actual question set, summary format, and escalation thresholds. The skill is only as opinionated as the reference files you hand it.
Test against a contract whose terms you know. Run on a contract whose key terms you can verify by hand — total value, renewal date, liability cap, termination rights. Confirm every line in the output cites a §reference and that anything ambiguous is rendered as Ambiguous — see legal rather than smoothed over.
Wire in the escalation triggers your GC cares about. The default trigger list is conservative; add your firm’s own (e.g. data-residency clauses for a regulated business, audit-rights thresholds for a public company).
Use on demand. Drop the contract file in, name the audience, attach the summary to the approval / handoff / renewal artifact.
What the skill actually does
SKILL.md runs five steps in order. The two-pass design — extract first, then filter — exists to keep the audience lens from causing the skill to skip a clause it would have judged irrelevant on first read.
Clause-by-clause extraction. Walks the contract top to bottom and captures every clause that affects parties, term, payment, renewal, termination, liability, indemnity, IP, confidentiality, data, change control, governing law. Verbatim or near-verbatim, with §references. Absent clauses are recorded explicitly — “no liability cap” is a finding, not silence.
Audience-aware filtering. Loads the named audience’s question set (exec, ops, or finance) from the question library and maps each question to the clauses that answer it. Unanswered questions render as Not addressed in contract, never as inferred industry defaults.
Structured summary. Renders the audience-matched template from references/2-summary-format-by-audience.md. Every line carries a §reference; nothing is rendered without one.
Ambiguity and escalation pass. Re-reads the rendered summary against references/3-escalate-to-legal-triggers.md. If a hard trigger fires (uncapped liability, MFN clause, unilateral termination for convenience, etc.), prepends the output with a COUNSEL REVIEW REQUIRED block naming the trigger and §. If a clause is genuinely ambiguous, replaces the line with Ambiguous — see legal plus the §reference.
Watch-outs. Renders the per-audience watch-outs, each paired with the specific clause that triggered it and the action the audience should take.
Cost reality
Token cost is the dominant variable, and it scales with contract length more than with audience choice.
Contract length
Approx. input tokens
Cost per summary (Claude Sonnet)
Cost per summary (Claude Opus)
Short (5-10 pages, MSA short-form)
~8k
~$0.04
~$0.20
Standard (15-30 pages, typical SaaS MSA)
~25k
~$0.10
~$0.50
Long (50-80 pages, enterprise MSA + exhibits)
~70k
~$0.25
~$1.20
Heavy (100-200 pages, M&A SPA, IP licensing)
~150k
~$0.55
~$2.60
Output is small (typically under 2k tokens) and rounds to noise against input cost.
At a typical mid-market legal-ops volume of 40 contracts per month (20 standard, 15 long, 5 heavy), monthly cost on Sonnet runs around $13-20; on Opus around $60-90. A team running every renewal pack through the skill at 200 contracts per month is in the $60-100/month range on Sonnet, $300-450 on Opus. Both bands are dwarfed by the salary cost of the half-hour-per-contract a paralegal would otherwise spend producing the same summary.
Success metric
Track two metrics. Both should move the same direction or you have a calibration problem.
Time-to-summary. Wall-clock from “contract landed” to “summary attached to approval workflow.” Baseline before this skill is typically 30-90 minutes (paralegal review). Target after adoption is under 10 minutes (skill run + reviewer spot-check).
Downstream legal escalation rate. Of the contracts run through the skill, what fraction reach counsel after the summary lands? This should go up when the skill is calibrated correctly — the escalation triggers exist to surface contracts that need legal eyes that previously got rubber-stamped because no one read the paper. If the rate goes down, the skill is over-paving and the escalation triggers in references/3-escalate-to-legal-triggers.md need to be tightened.
If time-to-summary drops and escalation rate rises, the skill is working. If only the first happens, you have built a confidence machine that is hiding risk.
vs alternatives
Ironclad executive summaries (CLM-native). Ironclad’s Workflow Designer can emit summaries from contracts that originated in Ironclad. The summary is constrained to data Ironclad already has structured. Use Ironclad if your contracts live in Ironclad and you need summaries inside that workflow. Use this skill if your contracts arrive as .docx/.pdf from outside the CLM, or if you need an audience cut Ironclad’s templates do not support.
Spellbook (Word add-in). Spellbook focuses on drafting and redlining inside Word and includes a summary capability. It is optimized for the lawyer-in-Word workflow. Use Spellbook if your legal team lives in Word and wants drafting and summary in one pane. Use this skill if your audience is non-lawyers and you want audience-tuned summaries that flag escalation triggers explicitly.
Manual paralegal abstracts. A trained paralegal produces the best summaries when given a clean exemplar to follow. Cost is the half-hour-per-contract problem and the inconsistency problem (two paralegals will emphasize different clauses). Use the skill as the first pass, the paralegal as the spot-check.
Watch-outs
Over-summarization losing nuance. A “12-month liability cap” reads simply but the carve-outs (gross negligence, IP indemnity, data-breach claims) often define the actual exposure. Guard: the skill always emits the cap line and a separate “Cap carve-outs” line citing the carve-out section verbatim; if no carve-outs are found, the line reads “No carve-outs found — verify with counsel.”
Audience mismatch. Surfacing renewal-pricing detail to an exec who only needs the sign-off gist creates noise; surfacing only the sign-off gist to finance leaves the budget model wrong. Guard: the audience parameter is required; the skill refuses to render without it and asks the caller to specify.
Citing the summary back as legal interpretation. Downstream readers sometimes treat the summary as the contract. Guard: every output is prefixed with Not legal advice and every line carries a §reference so the reader can verify against the source. If a reader forwards the summary as the basis of a legal position, the §refs make it self-evident that the contract — not the summary — is the source of truth.
Stale draft summarized as executed. Summaries written from redlines are wrong by the time the contract is signed. Guard: the skill detects draft markers (track changes, “DRAFT” watermark, version markers in the filename) and prepends a warning that the summary may not match the executed text; re-run after signature.
Stack
Pair with clause extraction for the structured-data version of the same contract, or with contract redline for the negotiation-stage workflow that runs before this one fires.
---
name: contract-summary
description: Produce an executive-readable summary of an executed contract for non-lawyer audiences (exec, ops, finance). Takes a contract file plus the audience type and returns a one-page Markdown summary that surfaces what each audience actually needs to act on, with explicit "ambiguous — see legal" fallbacks. Not for legal advice, dispute analysis, or privilege determination.
---
# Contract summary
## When to invoke
Whenever a non-lawyer needs to understand what is in an executed (or near-executed) contract well enough to do their job: an exec deciding whether to sign off, a CFO modeling cash flow, a deal team briefing the board, an ops owner who has just inherited responsibility for the counterparty relationship, a procurement lead preparing a renewal.
The output is a structured one-page Markdown summary tuned to the audience the caller names. It is a reading aid, not a substitute for the contract.
Do NOT invoke this skill for:
- Legal advice. The skill summarizes; it does not opine on whether a clause is enforceable, whether a position is defensible, or whether to sue. That is counsel's job.
- Dispute analysis. If the question is "did they breach?", read the contract with counsel against the specific facts. A summary will flatten the nuance the dispute turns on.
- Privilege determination. The skill cannot decide whether a document is protected. Privilege depends on facts the contract does not contain (who saw it, in what capacity, for what purpose).
- Non-Tier-A AI vendors. Run this on Claude only. Do not pipe contracts to consumer-grade LLMs whose data-handling posture you have not reviewed; contracts contain commercial terms, counterparty PII, and sometimes regulated data.
## Inputs
- Required: `contract` — the executed contract as `.docx`, `.pdf`, or pasted Markdown. If it is a redline or draft, the skill prepends a warning to the output that the summary may be stale.
- Required: `audience` — one of `exec`, `ops`, `finance`. Drives which questions the watch-outs section answers and which clauses get emphasis. See `references/2-summary-format-by-audience.md` for the per-audience question library.
- Optional: `template` — path to a firm-specific summary template that overrides the default output format (e.g. your GC's preferred structure for board-pack approvals).
- Optional: `prior_version` — a previous executed version of the same contract. When provided, the skill adds a "What changed since the prior version" section instead of the standard intro.
## Reference files
Always load the following from `references/` before summarizing. Without them, the skill falls back to a generic format and misses the audience-specific questions that make the summary actionable.
- `references/1-audience-question-library.md` — the canonical questions exec, ops, and finance audiences ask of a contract. The skill uses this to decide which clauses to surface and how to phrase the watch-outs.
- `references/2-summary-format-by-audience.md` — the literal Markdown scaffold for each audience. Copy the matching template, fill it in from the contract, leave any unanswered field as `Ambiguous — see legal`.
- `references/3-escalate-to-legal-triggers.md` — the list of clause patterns and ambiguity signals that force the skill to stop summarizing and emit an escalation block. When a trigger fires, the skill returns the partial summary plus an explicit "Counsel review required before acting on this summary" header.
## Method
Run these five sub-tasks in order. The skill is single-pass per section but two-pass overall: extraction first, then audience-aware filtering.
### 1. Clause-by-clause extraction
Walk the contract top to bottom and extract a structured list of every clause that affects: parties, term, payment, renewal, termination, liability, indemnity, IP, confidentiality, data handling, change control, governing law. Do not summarize yet. Capture the clause verbatim or near-verbatim, with section reference (e.g. "§7.2"), so the next pass has citations to point at.
If a clause type is absent, record "not present in contract" rather than inferring its content from boilerplate. Absent clauses are often the most important watch-out (e.g. no liability cap, no termination for convenience).
### 2. Audience-aware filtering
Load the question set for the named `audience` from `references/1-audience-question-library.md`. For each question, find the clause(s) from step 1 that answer it. If no clause answers a question, mark it `Not addressed in contract` — do not invent an answer from "industry standard" defaults.
This step is the reason the skill is two-pass: extracting first and filtering second prevents the audience lens from causing the skill to overlook a clause it would have judged irrelevant on first read.
### 3. Structured summary
Render the audience-matched template from `references/2-summary-format-by-audience.md`. Quote section references in every line so the reader can jump back to the source. Use plain language; never use legalese the audience would have to translate.
### 4. Ambiguity and escalation pass
Re-read the rendered summary against `references/3-escalate-to-legal-triggers.md`. For each trigger:
- If a trigger fires (e.g. uncapped indemnity, MFN clause, unilateral termination for convenience by counterparty only, ambiguous governing-law fork), prepend the output with a `Counsel review required` block naming the trigger and the §.
- If a clause is genuinely ambiguous (drafting error, contradicting sections, undefined defined term), replace the rendered line with `Ambiguous — see legal` plus the § reference. Never paper over ambiguity with a confident paraphrase.
### 5. Watch-outs section
Render the per-audience watch-outs from the question library. Each watch-out is paired with the specific clause that triggered it and the action the audience should take (read §X, ask counsel about Y, model scenario Z in the budget).
## Output format
Render exactly this structure (substitute the audience block for the named audience). Anything the contract does not answer is rendered as `Ambiguous — see legal` or `Not addressed in contract`, never elided.
```markdown
# Contract summary — {Counterparty} / {Contract type} ({Effective date})
> Audience: {exec | ops | finance}
> Source: {filename}, {N} pages, executed {date}
> Prepared by: contract-summary skill (Claude). Not legal advice.
## Parties
- Customer: {legal entity, jurisdiction} (§{ref})
- Vendor: {legal entity, jurisdiction} (§{ref})
- Affiliates in scope: {list or "none"} (§{ref})
## Term
- Effective date: {date} (§{ref})
- Initial term: {duration} (§{ref})
- Initial term ends: {date}
- Renewal mechanism: {auto-renew / opt-in / none} (§{ref})
- Renewal notice deadline: {date or "n/a"}
## Key obligations
- Our side: {plain-language obligations, with §refs}
- Their side: {plain-language obligations, with §refs}
- Service levels / acceptance criteria: {summary or "Not addressed"} (§{ref})
## Key money flows
- Total contract value: {amount, currency} (§{ref})
- Payment cadence: {monthly / quarterly / annual / milestone} (§{ref})
- Payment terms: {Net X} (§{ref})
- Price escalators: {CPI / fixed / none} (§{ref})
- Late payment penalties: {summary or "Not addressed"} (§{ref})
- Taxes: {who pays} (§{ref})
## Renewal terms
- Auto-renewal: {yes / no} (§{ref})
- Renewal length: {duration}
- Notice to prevent renewal: {duration} before term end
- Pricing on renewal: {locked / market / capped escalator} (§{ref})
## Termination rights
- Termination for convenience — us: {yes / no, notice} (§{ref})
- Termination for convenience — them: {yes / no, notice} (§{ref})
- Termination for cause: {triggers, cure period} (§{ref})
- Effect of termination: {data return, transition assistance} (§{ref})
## Watch-outs for {audience}
- **{Watch-out 1}** — {one-sentence why this matters to this audience}.
Action: {read §X / ask counsel about Y / model scenario Z}.
- **{Watch-out 2}** — ...
- **{Watch-out 3}** — ...
```
If a `Counsel review required` trigger fires, the entire output is prepended with:
```markdown
> COUNSEL REVIEW REQUIRED before acting on this summary.
> Trigger(s): {list}, see §{refs}.
```
## Watch-outs
- **Over-summarization losing nuance.** A "12-month liability cap" reads simply but the carve-outs (gross negligence, IP indemnity, data-breach claims) often define the actual exposure. Guard: the skill always emits the cap line *and* a separate "Cap carve-outs" line citing the carve-out section verbatim; if no carve-outs are found, the line reads "No carve-outs found — verify with counsel."
- **Audience mismatch.** Surfacing renewal-pricing detail to an exec who only needs the sign-off gist creates noise; surfacing only the sign-off gist to finance leaves the budget model wrong. Guard: the audience parameter is required; if absent, the skill refuses to render and asks the caller to specify.
- **Citing the summary back as legal interpretation.** Downstream readers sometimes treat the summary as the contract. Guard: every output is prefixed with `Not legal advice` and every line carries a §reference so the reader can verify against the source. If the caller forwards the summary as the basis of a legal position, the reference back to the §s in the contract is built into the document.
- **Stale draft summarized as executed.** Summaries written from redlines are wrong by the time the contract is signed. Guard: the skill detects draft markers (track changes, "DRAFT" watermark, "v" in filename) and prepends a warning that the summary may not match the executed text; re-run after signature.
# Audience question library — TEMPLATE
> The contract-summary skill loads this file before filtering the
> extracted clause list. Each audience has a different set of questions
> they ask of a contract; the summary should answer those questions
> and only those questions. Replace the placeholder questions below
> with the ones your team actually asks — the defaults are a starting
> point, not your firm's voice.
## Exec audience
The exec is approving sign-off. They want the gist plus anything that would embarrass them in a board meeting six months from now.
Standard questions:
1. What does this commit us to do, in one sentence?
2. What does it commit them to do, in one sentence?
3. What is the total contract value over the initial term?
4. What is the longest we can be locked in (initial term + worst-case renewals)?
5. Can either side walk away for convenience? With what notice?
6. Is there anything unusual that a reasonable peer would push back on? (Uncapped liability, MFN clause, exclusivity, IP transfer, change-of-control restriction.)
7. Are there obligations on us that require operational changes (audit rights, security controls, named personnel) that the business has not yet planned for?
Watch-outs to surface for exec:
- Auto-renewals with short notice windows (>12 month renewal, <90 day notice — high regret risk)
- Change-of-control restrictions (relevant to any future M&A conversation)
- Exclusivity, non-solicit, or non-compete obligations
- Personal guarantees or executive sign-off requirements
- Any clause that names an individual (key-person dependency)
## Ops audience
The ops audience is responsible for executing the contract day-to-day: delivering against SLAs, tracking deliverables, managing the relationship.
Standard questions:
1. What are we obligated to deliver, by when?
2. What service levels apply, and what are the consequences of missing them?
3. What deliverables, reports, or notices do we owe them on a recurring basis?
4. What deliverables, reports, or notices do they owe us?
5. Who is the named contact / point of escalation on each side?
6. What change-control process governs scope changes?
7. What are the operational triggers for renewal notice (date, responsible party, system of record)?
Watch-outs to surface for ops:
- Service-level commitments without a clear measurement methodology
- Reporting cadences that are not yet wired into the team's calendar
- Notice obligations that require formal written delivery (vs email)
- Audit cooperation obligations (who handles, on what timeline)
- Data return / destruction obligations on termination
## Finance audience
The finance audience is modeling cash flow, recognizing revenue or expense, and forecasting renewal economics.
Standard questions:
1. What is the total contract value, by year of the initial term?
2. What is the payment cadence and net terms?
3. Is the price fixed, escalating, or usage-variable? If escalating, on what basis (CPI, fixed %, market)?
4. Are there minimum commitments or take-or-pay provisions?
5. Who pays which taxes (sales, VAT, withholding)?
6. What are the consequences of late payment (interest rate, suspension, termination)?
7. What is the renewal pricing structure, and when does notice need to be given to avoid an unfavorable renewal?
Watch-outs to surface for finance:
- CPI escalators without a cap (open-ended budget exposure)
- Auto-renewal with locked-in price increases
- Currency exposure (counterparty in different currency, no FX clause)
- Withholding tax obligations the AP team has not modeled
- Termination liability (early-termination fees, payment for remaining term)
## Last edited
{YYYY-MM-DD}
# Escalate-to-legal triggers — TEMPLATE
> When the contract-summary skill detects any of the patterns below,
> it stops summarizing the affected line and emits a
> `Counsel review required` block at the top of the output, naming
> the trigger and the §reference. The skill never paraphrases over
> these patterns; over-summarization here is the highest-risk failure
> mode. Replace the placeholder thresholds with your firm's actual
> escalation criteria — the defaults are a conservative starting
> point.
## Hard triggers (always escalate)
| Pattern | Why it triggers | What the summary emits |
|---|---|---|
| Uncapped liability or "no cap on indemnity for X" | Unbounded exposure | "COUNSEL REVIEW REQUIRED — uncapped liability at §{ref}" |
| Liability cap < 6 months of fees | Unusually low cap | Block + cap stated verbatim |
| Mutual indemnity replaced with one-way (against us) | Asymmetric risk | Block + clause stated verbatim |
| MFN ("most favored nation") clause | Constrains future deals | Block + §ref |
| Unilateral termination for convenience (counterparty only) | We can be exited at will | Block + notice period |
| Change of control without consent | Blocks M&A | Block + §ref |
| Exclusivity or non-compete on us | Limits future business | Block + scope and duration |
| IP assignment beyond deliverables | Over-broad IP transfer | Block + scope |
| Personal guarantee | Individual on the hook | Block + §ref |
| Choice-of-law in counterparty-favorable jurisdiction without arbitration | Litigation venue risk | Block + jurisdiction |
| Auto-renewal > 12 months with notice window < 90 days | High regret risk | Block + the dates |
## Soft triggers (escalate if combined)
Any single one of these does not force an escalation, but two or more in the same contract trigger a `Counsel review recommended` (lower severity than `required`):
- Termination for cause without a cure period
- "Sole discretion" wording in an obligation that affects us
- Defined terms used before they are defined
- References to schedules / exhibits that are not attached
- Currency or governing law mentioned inconsistently across sections
- Notice obligations requiring physical mail or specific carriers
- Audit rights without scope or frequency limits
- Data-handling obligations citing standards not named in the contract (e.g. "industry-standard practices" with no definition)
## Ambiguity triggers (replace line with "Ambiguous — see legal")
The skill does not paraphrase past these — it leaves the gap visible so legal review fills it in:
- Two sections appear to conflict on the same point (e.g. term ends on different dates in different clauses)
- A defined term is used but never defined
- A monetary amount is given in two different units or currencies
- Dates are inconsistent (e.g. effective date later than first payment date)
- A signature block is missing or incomplete
- The contract refers to a counterparty entity name that does not match the parties block
## Output format for an escalation block
```markdown
> COUNSEL REVIEW REQUIRED before acting on this summary.
> Trigger(s):
> - {Trigger name} (§{ref}): {one-line description}
> - {Trigger name} (§{ref}): {one-line description}
> Do not rely on any line of this summary as legal interpretation.
> Counsel should review the cited sections before sign-off.
```
## Last edited
{YYYY-MM-DD}