From Prompts to Context Engineering

PLA 2872Activity 5·9 min read

Why it Matters

You upload a lease and type “identify any risks in this contract,” and the model returns a few observations about rent escalation and the security deposit while missing the personal guarantee buried in the miscellaneous section and the absence of any hurricane clause on a property in Coral Gables. It missed them because you let it decide what mattered. Professional AI use has moved from clever prompts to designed context: reusable skill documents that encode a method, uploaded sources the model must cite, and instructions that fix the categories and the output before the model reads a word. This activity has you audit the same lease three ways, measure the difference, and then write your own skill document for a different document type.

Current Context

In 2026 NetDocuments published a Legal Context Engineering Benchmark Report that held the model constant and varied only the context layer across 300 questions drawn from ten real matters and 874 documents, finding that structured context produced answers about 7 percent more accurate at 18 percent lower cost and proposing cost per correct answer as the metric firms should use to judge their AI investments. The technique has a standard form: in October 2025 Anthropic published the Agent Skills format, a folder with a SKILL.md file that a model loads only when a task calls for it, and the Model Context Protocol gives models a common way to reach documents and tools. The lease audit you have been running as a classroom exercise is now the way legal AI products are built, and the SKILL document you write in Part 4 uses the same structure.

Key Concepts

Prompt and Context

A prompt is the request; context is everything the model can see when it answers, including instructions, uploaded documents, examples, and prior turns. The quality of an answer tracks the quality of the context far more than the cleverness of the request.

System and Project Instructions

Standing instructions that apply to every conversation in a project or custom assistant, such as the role, the output format, and the rules for citing sources. They are how a firm makes an AI tool behave the same way for every user.

Skill Document

A written methodology the model follows for a task: purpose, when to use it, the steps in order, the categories to check, the rating scale, and the output format. A good SKILL turns a general model into a specialist for one job, and it can be reviewed and improved like any other work product.

Grounding

Supplying the authoritative text, such as the lease itself and the governing statute, so the model quotes what is in front of it instead of recalling what it once read. Grounded output can be verified; ungrounded output has to be re-researched.

Model Context Protocol

An open standard that lets an AI tool connect to document systems, databases, and other software through a common interface. It is how a model in a law firm reaches the document management system, and it is why context engineering now includes deciding which tools a model may use.

Evaluation Rubric

A fixed set of criteria for scoring an AI output, such as coverage of required categories, accuracy of section citations, correct risk ratings, and identification of missing provisions. Without a rubric, comparing two audits is a matter of taste.

Florida-Specific Provisions

Lease terms Florida practice requires a reviewer to check: force majeure language for hurricanes and named storms, governing law and venue, and the unenforceability of confession of judgment clauses. A model that knows contract law in general will miss them unless the context tells it to look.

Resources

Sample Lease Miracle Mile (PDF) (The commercial retail lease for a Coral Gables cafe that you audit; the version for reading.)
Sample Lease Miracle Mile (Markdown) (The same lease as plain text, which AI tools read more reliably than PDF.)
SKILL: Commercial Lease Auditor (PDF) (The skill document that encodes the audit methodology; read it before Part 2.)
SKILL: Commercial Lease Auditor (Markdown) (The same SKILL as plain text for uploading.)
Florida Statutes Chapter 83, Landlord and Tenant (Part I governs nonresidential tenancies; paste it as grounding in the third audit and use it to verify findings.)
Landlord-tenant law, Cornell Wex (A plain-language overview of the doctrine behind the audit categories.)
Equipping agents for the real world with Agent Skills, Anthropic (The October 2025 post that defines the SKILL.md format and progressive disclosure; the model for Part 4.)
Agent Skills overview, Anthropic documentation (The specification for a skill folder and its metadata.)
Effective context engineering for AI agents, Anthropic (Why what is in the context window matters more than the prompt.)
Model Context Protocol (The open standard for connecting models to documents and tools.)
2026 Legal Context Engineering Benchmark Report, NetDocuments (The measurement behind Current Context; read the methodology section.)
NotebookLM (A source-grounded tool you can use for the third audit by loading the lease and the statute as sources.)

What to Do

In this activity you audit one commercial lease three ways, compare the results with a rubric, verify the findings against the lease and the statute, and then write and test a skill document of your own. You post the comparison, the verification results, and your SKILL.

Part 1: Read the Lease and the SKILL

Read the Miracle Mile lease once so you know the document before any machine touches it, and read the Commercial Lease Auditor SKILL to understand the method it imposes: the categories, the risk ratings, and the required output. Skim Part I of Chapter 83 so you know what Florida law says about nonresidential tenancies.

Part 2: Run Three Audits

For the naive audit, open a fresh conversation, upload the lease, and give only this instruction: “Review this lease and identify any issues or concerns.” Accept the output as is. For the SKILL audit, open a new conversation, upload the SKILL document first, then the lease, and instruct the model to conduct a lease audit following the SKILL’s methodology. For the grounded audit, open a third conversation, upload the SKILL, the lease, and the text of Chapter 83 Part I, and instruct the model to follow the SKILL and to cite the statute section for every Florida-specific finding; NotebookLM with the three documents loaded as sources works well here. Save all three outputs.

Part 3: Compare and Verify

Use the Audit Comparator to score the three audits against a rubric and to explain the differences. Then choose three findings from the grounded audit and use the Finding Verifier to check each against the actual lease language and, where relevant, the statute: does the section say what the model claims, is the risk rating right, and is the legal conclusion correct? Finally, read the lease yourself for at least two issues none of the audits handled well, because the point of the exercise is that structured context narrows the gap without closing it. When you have your results, use the Context Engineering Explainer to make sense of why the three audits differed and where even the grounded one fell short.

Part 4: Write Your Own SKILL

Choose a different document type: a Florida residential lease, an attorney engagement letter, or a demand letter. Use the SKILL Architect to draft a skill document with the same parts as the Commercial Lease Auditor, including purpose, when to use it, the method in order, the categories to check, a rating scale, the output format, and the Florida-specific checks. Test it once on a sample document of that type and note what it caught and what it missed.

Part 5: Post Your Deliverables

Create a single new discussion thread containing a comparison of about 400 words across the three audits, your verification results for the three findings and the two issues you found yourself, your SKILL document, and a short note on how your SKILL performed in its first test and what you would change.

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Suggested AI Prompts

Use these prompts as a starting point, then adjust them to fit your goal. Strong prompting develops through trial, revision, and testing. It’s a foundational skill that grows into more advanced AI work such as context engineering and agent-based workflows.

Audit Comparator

Act as a supervising attorney evaluating three lease audits prepared by a paralegal: one from a simple prompt, one following a structured SKILL document, and one following the SKILL with the governing Florida statute supplied as grounding. Score each audit from 1 to 5 on these criteria: coverage of the SKILL’s required categories, consistency of risk ratings, specificity of section citations, identification of missing provisions rather than only problematic language, and Florida-specific issues caught. Present the scores in a table, then explain the differences by pointing to particular findings or gaps in each audit. Tell me which audit you would trust most and why. Wait for me to paste the three audits.

A rubric applied the same way three times. Fixed criteria and a numeric scale turn a comparison that could be a matter of taste into a measurement, and asking for specific findings behind each score makes the reasoning checkable.

Finding Verifier

Act as a skeptical legal reviewer checking AI work product. I will paste one finding from a lease audit along with the actual lease language it references and, where relevant, the text of the Florida statute. Verify whether the lease section says what the finding claims, whether the risk rating fits what the text shows, and whether the legal conclusion is correct under the statute I supplied. Quote the lease or statute language that supports or undercuts the finding. Be critical; I need to know whether this finding can go to an attorney. Wait for me to paste the finding and the source text.

Verification against supplied text. Handing the model the lease section and the statute confines it to reading and comparing, which is where it is reliable, and the quotation requirement leaves you a trail to check.

SKILL Architect

Act as an experienced legal technologist who writes skill documents for AI tools. I want to build a SKILL document for reviewing a specific type of legal document, which I will name. Interview me briefly about the purpose, the users, and the risks that matter most. Then draft a SKILL with these parts: a one-paragraph purpose and when to use it, the method as ordered steps, the categories the review must cover, a rating scale with definitions, the required output format, the Florida-specific checks, and a short list of things the reviewer must verify by hand rather than trust. Keep the instructions concrete enough that a different model following them would produce the same kind of output. After drafting, ask me to test it on a sample document and offer to revise based on what it missed.

Building the context instead of a prompt. The interview surfaces what matters before drafting, the fixed parts mirror the format the industry has adopted, and the test-and-revise loop treats the SKILL as work product to be improved rather than a one-time request.

Context Engineering Explainer

Act as a legal technology consultant helping a paralegal understand why structured context changes AI output. I just compared a naive prompt, a SKILL-guided audit, and a grounded audit of the same lease, and I will describe the differences I saw. Explain what was happening inside the model in each case, why supplying the statute changed the findings, what a good SKILL document does that a prompt cannot, and where even well-designed context still fails. Then help me think about which other legal document tasks the approach would fit and which it would not. Ask me to describe my results to begin.

Reflection tied to your own results. Explaining the mechanism after you have seen the difference makes the concept stick, and asking where the approach fails keeps the lesson from becoming a sales pitch.