Context, Not Prompts

What actually changed the answer?

You give a model a commercial lease and ask it to find the risks, and it returns a decent list. You give the same model the same lease and a sharper request, and the list barely moves. Then you give it the same lease, a written method to follow, and the governing statute, and the list changes. It catches a personal guarantee buried in the miscellaneous clauses, and it notices that the lease has no hurricane clause at all, a gap the first two runs passed over. What changed?

The wording of the request did not change the answer. What changed was the context: everything the model could see when it answered. The claim this reading defends is blunt: the prompt is not the unit of work. The unit is the context you assemble before the model reads a word, and one of the companies that build these tools now says the same about its own products (opens in a new tab). The activity has you test the claim, auditing one lease three ways and scoring what comes back, and this reading walks the argument behind the experiment.

A method beats a better sentence

Reword a request all you like and the model still decides what matters. A lease hides its worst terms in dull places. The personal guarantee that makes a café owner answer personally for the company’s rent sits in the miscellaneous clauses of the Miracle Mile lease, at section 15.6, not in a headline section near the top. Ask a general question and a general model may sail right past it.

The first layer of context that fixes this is a written method. A skill document tells the model what to check and how to report it. The Commercial Lease Auditor you use here fixes ten categories, four risk ratings, a Florida checklist, and one report format. Now the model is not deciding what a lease review is; it is running one. It has to walk every category, so the guarantee turns up under the first heading, which asks in plain words whether a guarantor provision is present. The same method catches the terms a reader skims past: an eighteen percent annual interest charge on late rent, an automatic renewal a tenant loses by missing a fifteen-day window.

This is the move a firm makes with system and project instructions, the standing rules that ride along in every conversation. A skill document is that method written down once and reused. Anthropic’s Agent Skills format (opens in a new tab) is one public standard for it: a file the model loads only when a task calls for it. The method travels with the work instead of living in one person’s head. Whoever writes that file decides what every later audit will notice.

Put the source in the room

A method tells the model what to look for. It does not tell the model what the law is, and there the model is still working from memory. Ask it about a Florida commercial lease and it may reach for the state’s residential security-deposit statute, section 83.49, Florida Statutes, because that is the rule it has seen most often. That statute governs apartments, not storefronts. A commercial deposit is a matter of contract, and Part I of chapter 83 (opens in a new tab), which covers nonresidential tenancies, carries no deposit section at all.

Grounding closes that gap. Put the lease and the statute in the window and tell the model to cite a section for each Florida finding. Then it quotes what is in front of it instead of recalling what it read in training. Now the model can see three things: Part I is thin, a commercial deposit answers to the lease and not to a statute, and the lease has no force-majeure clause at all. The auditor’s Florida checklist expects one, naming hurricanes and named storms, for a storefront in Coral Gables. It flags the confession-of-judgment clause in section 10.3 the same way, as a provision for a Florida attorney to weigh. A missing provision is the hardest kind to catch, because there is no language on the page to react to, and grounding paired with a checklist is what surfaces it.

Building the context, and proving it worked

In a real firm the lease and the statute do not arrive by copy-paste. They live in a document management system, and the model reaches them through plumbing. The Model Context Protocol is an open standard for that connection, a common way for a tool to pull the right file, matter, or database into the window. At that scale, deciding what the model sees includes deciding which sources it may touch and when, and the work becomes an engineering job rather than a knack for phrasing.

That raises a fair question: how do you know better context produced a better answer, and not just a longer one? You measure. An evaluation rubric is a fixed scorecard, applied the same way to every output, so a comparison stops being a matter of taste. Benchmarks work this way. NetDocuments’ 2026 benchmark study (opens in a new tab) held the model constant and changed only what the model was given, then scored the answers for accuracy and cost. It proposed cost per correct answer as the number a firm should watch. In the activity you run that experiment by hand: one lease, three versions of what the model is given, five scored lines.

Where context stops

The proposition has a limit, and being honest about it is part of the skill. Better context narrows the gap between what a model catches and what a careful lawyer catches; it does not close it. A grounded model can still misread a clause it quotes correctly, rate a real risk too low, or miss a term the checklist never thought to name. That is why the activity ends with you, not the machine. You verify two of the grounded audit’s findings against the lease and the statute yourself, reading the words and quoting the ones that settle it. The auditor’s own disclaimer says the same thing in one line: every finding goes to a licensed attorney before anyone relies on it.

So the person who designs the context decides what the machine notices, and the person who checks it decides what the client can trust. Those two jobs, designing and checking, outrank clever phrasing, and a paralegal can hold both. In the activity you run the three audits, score them yourself before the Audit Comparator does, reconcile the two sets of scores, and explain what the context, and only the context, changed. Start with Part 1: read the lease before a machine touches it.