Business Β· Guide β0002 Β· 4 min read
Never ask if it's "fine."
Triage any vendor contract, lease, or MSA in about 20 minutes β free, in Claude's web app, no card β by making the model transcribe the eight clauses that cost money instead of asking its opinion. The catches, stated up front: the free tier's message caps are unpublished and vary by model and hour, your conversations can be used for model improvement, and this is triage that decides when you need a lawyer β not whether.
Numbers wearing an "as of" chip drift. The eight-clause framework doesn't.
The centerpiece β Red-Flag Triage
Score the contract before you sign it
Answer eight questions from the extraction table (Step 2 below gives you that table in four minutes). The verdict updates live and tells you which branch you're on: sign it, renegotiate it, or hand it to a lawyer. Nothing is sent anywhere β this runs entirely inside the page.
score
Triage is a ranking of what to escalate β not legal advice. Regulated industries, employment agreements, and IP-heavy deals start at RED regardless of score.
The cost of one "looks standard"
$14,400One auto-renewal clause on page 51, missed because the question was "any red flags?" instead of "transcribe clause 2."
01 β Why extraction wins
A clerk, not a judge
Models trained on the internet's collective hedging produce "fairly standard agreement" the way a coin produces heads β it's the statistically safe completion of a vague question, not a judgment about your document. Ask for a verdict and you get the average of the internet. Ask for transcription and the same model becomes a tireless clerk that quotes page 51 verbatim.
That's why the protocol has exactly eight clauses and an ABSENT-or-quote rule. Eight is what fits in one readable table and what covers the clauses that actually cost money: renewal mechanics, exit rights, liability, indemnity, IP, fees, and where disputes happen. Everything else is prose.
And it's why the division of labor is fixed: the model reads; you judge. The model can quote a 90-day notice window, but only you know whether anyone at your company will remember it in January. That judgment against your own calendar β not the model's opinion β is what the triage calculator above turns into a verdict.
When a contract does deserve a lawyer, the 8-row table is what you hand over. You're paying for judgment, not page-turning β and the table cuts the billable read from hours to minutes.
What it replaces
Tick what you'd otherwise pay for first-pass review β then run the swap.
Pricing is typical-range, secondary-source β treat as an estimate, not a quote.
02 β The walkthrough
Five steps, one verdict
Anonymize before you upload
Open the contract and find-replace: party names become "Customer" and "Vendor", dollar amounts become rounded placeholders, everything else stays. The free tier's terms allow conversations to be used for model improvement β there is no un-upload, so redaction happens on your copy before it ever touches a browser tab.
Anonymizing takes four minutes. Skipping it means your real vendor names and rates live in a training corpus forever β and a confidentiality clause you're about to sign probably forbids exactly that.
The extraction pass
In Claude, pick a mid-tier model, not the flagship. Transcription is a clerk's job, and mid-tier models get you more messages per day β burning the flagship's small budget on page-turning is how people run dry after one contract. Upload the anonymized PDF and send:
You are a contract-transcription assistant, not an advisor. From the attached agreement, extract the following eight items. For each: quote the contract VERBATIM, give the section or page reference, and if the item does not appear, write exactly ABSENT. Do not evaluate, summarize, or reassure. 1. Initial term and renewal term 2. Auto-renewal notice window β days, and the required delivery channel 3. Termination for convenience β allowed or not, and any exit fee 4. Liability cap β the multiple of fees, and any carve-outs 5. Indemnification β direction (who indemnifies whom) and scope 6. IP ownership, and any license either party grants the other 7. Fees β escalators, minimum commitments, late penalties, true-ups 8. Governing law and dispute path Output as an 8-row table. If you cannot find something, say ABSENT β never guess a clause.
I ran the naive version first: uploaded a 62-page MSA and asked, "Any red flags?" The reply opened "Overall, this looks like a fairly standard agreement," flagged two cosmetic items, and moved on. Nine months later we missed a renewal clause requiring 90 days' written notice β delivered by certified mail to a Delaware address β and auto-renewed for a full year at $1,200 a month for a tool the team had stopped using in October. $14,400. The clause was on page 51. The extraction pass takes four minutes and would have quoted it verbatim in row two of the table.
Spot-check three quotes, then add your numbers
Ctrl-F three of the verbatim quotes in the original PDF. If a cited section doesn't exist, the model invented it β re-run with "quote only what you can point to." Ninety seconds of checking is the only firewall between you and a hallucinated clause number; on contracts over 40 pages, that's where the mistakes hide.
Then fill in what the model can never know: your annual spend against the cap multiple, the notice window against your actual calendar, the escalator against your actual budget. The model reads; you judge.
Failure scenarios, narrowly
Take the one to three worst rows from the triage calculator and only those. Scope discipline is the budget: broad questions get the internet's average answer; narrow questions about quoted text get specific ones β and on a message-capped free tier, that difference decides whether you finish the contract today.
Here is a clause quoted verbatim from our agreement: [paste the row's quote]. Without reassurance or hedging, list the three most expensive failure scenarios this clause permits β each with the trigger, who pays, and an order-of-magnitude cost for a business of our size ([employees / annual revenue]). Do not comment on whether the clause is "standard."
Verdict and calendar
Run the Red-Flag Triage calculator at the top of this page with your eight answers. GREEN: sign after your own read β and put the renewal date in two calendars before signing, not the week it matters. YELLOW: send the ask-script below; one email round-trip. RED: lawyer β hand them the table, not the PDF.
Draft a 120-word email to [vendor] requesting: [the specific change]. Reference the clause by section number. Neutral, professional, assumes the change is routine. No apologetic language.
Send it yourself. A letter from counsel changes the temperature of the conversation β and the price of the next one.
03 β Capacity math
Will the contract fit?
Token estimate: words Γ ~1.35 tokens/word. The 200K-token context window is the one hard, published number in this workflowas of 2025-06-19 β the free tier's message caps are deliberately unpublished and vary by model and demand.
What breaks
Never ask an LLM whether a contract is fine. "Fairly standard" is the statistically safe completion of a vague question, not a judgment about your document. Ask for transcription and the same model becomes a tireless clerk. The failure in the $14,400 story wasn't the tool β it was asking the tool for an opinion.
- Message caps are unpublished. Anthropic doesn't state the free tier's daily message count β it varies by model and demand, and the countdown appears only in the UI. The protocol survives it: extraction is idempotent, so if you run out mid-contract, re-upload in a fresh chat and re-run the same prompt. Nothing depends on conversation memory.
- Free-tier data usage. Conversations may be used for model improvement per Anthropic's consumer terms. That's why Step 1 exists β anonymize before upload, and check your account's privacy settings before sending anything sensitive.
- Hallucinated citations. The model may cite a Β§8.2 that doesn't exist. The three-quote Ctrl-F check in Step 3 is non-negotiable for exactly this reason.
- Scan-only PDFs. If the contract is a scan without a text layer, extraction gets garbage. Test by copy-pasting one paragraph from the PDF first; if nothing pastes, OCR it before starting.
- Giant contracts. A 200K-token windowas of 2025-06-19 fits roughly 250β300 pages, but accuracy degrades as you approach the limit. Over ~200 pages, split by exhibit, run the extraction prompt per chunk, and merge the tables yourself.
- The flagship temptation. On small message budgets, people pick the biggest model and run dry after one contract. Use a mid-tier model for extraction β it's a transcription job β and save the flagship for the failure-scenario pass.
- This is triage, not review. Regulated industries, employment agreements, and IP-heavy deals start at RED regardless of the score. The calculator ranks what to escalate; it doesn't replace the person you escalate to.
05 β Three things worth keeping
If you forget everything else
Never ask if a contract is fine β "fairly standard" is the safe answer to a vague question, and it costs $14,400 when it's wrong. Ask for verbatim transcription and the model does the one thing it's genuinely tireless at.
The ABSENT-or-quote rule is the hallucination firewall. A model forced to quote can be checked with Ctrl-F in ninety seconds; a model asked to summarize can quietly invent a clause number you'll never verify.
The model can quote a 90-day notice window, but only you know whether anyone at your company will remember it in January. Numbers against your calendar, cap multiples against your spend β that judgment is what the verdict actually runs on.
06 β Try it now
Start with the one you're already inside
- Pick a contract you've already signed β the one whose renewal date you can't name from memory.
- Anonymize it (Step 1) and run the extraction prompt (Step 2).
- Feed the eight answers into the Red-Flag Triage calculator above.
- YELLOW or worse? The renegotiation ask-script works mid-term too β windows re-open at every renewal.
- Calendar the notice date in two systems before you close the tab.