The CRAFT Prompting Framework for Legal Work

Why Most Legal Prompts Fail

Prompt quality strongly shapes output quality. When attorneys find AI output unusable for real legal work, a vague prompt is often part of the problem, though not all of it: models still have real limits, such as fabricated citations, which is why the verification step below matters. "Summarize this contract" tends to produce a generic overview; the same document with a structured prompt can produce a risk matrix that gives partner review a solid starting point. CRAFT provides the repeatable five-step method, and the advanced techniques below sharpen it further.

What Is the CRAFT Prompting Framework?

CRAFT stands for Context, Role, Ask, Format, Tone. Each element fixes a specific prompting failure:

• Context grounds the task in jurisdiction, practice area, and facts.
• Role tells Claude whose perspective to take, focusing its behavior and tone on your use case.
• Ask specifies the exact deliverable, preventing scope creep.
• Format structures the output for immediate use: memo, table, checklist, or redline.
• Tone sets the register so output matches its audience.

Step 1: Context, Ground the Task

Claude cannot infer your jurisdiction, governing law, or client situation. State them. A strong context includes the jurisdiction (e.g., "Delaware corporate law"), the matter type (e.g., "Series B preferred stock financing"), and constraints (e.g., "existing SAFE notes converting in this round").

Example: "We represent a Delaware-incorporated SaaS startup in a $15M Series B round. The lead investor requires a full-ratchet anti-dilution provision. California governs the side letter."

Step 2: Role, Set the Perspective

Telling Claude to "act as a senior corporate attorney specializing in venture capital transactions" focuses its behavior and tone on your use case, as Anthropic's prompting guide describes it. A short role is enough: the same guide notes that even a single sentence makes a difference, so there is no need to stack up credentials such as years of experience. As the note at the top of this post explains, describing who will read the output is often a better use of this step than an elaborate persona.

Step 3: Ask, Specify the Deliverable

State exactly what you need. Not "review this contract" but "identify every provision in this SPA that allocates indemnification risk to the buyer, flag deviations from ABA model terms, and recommend redline language for each." Define the deliverable, not the topic: "tell me about employment discrimination law" produces a textbook, while "draft a 3-page research memo analyzing whether our client's termination constitutes disparate treatment under Title VII, applying the McDonnell Douglas framework" produces work product.

Step 4: Format, Structure the Output

Legal work demands structure. Tell Claude: formal memo with IRAC analysis, bullet-point risk summary, comparison table, redline draft, or client-ready email. Specify section headers and ordering (e.g., highest risk first).

Step 5: Tone, Match the Audience

A federal court brief and a client advisory email need different language. Tell Claude the audience: "persuasive tone for a motion to dismiss," "plain English for a non-lawyer board member," or "analytical tone for an internal research memo."

CRAFT in Action: A Complete Example

A full CRAFT prompt for AI-assisted contract review:

"Context: We represent the buyer in a $50M asset purchase of a manufacturing business in Ohio. The seller has provided a 60-page APA with Ohio choice-of-law. Role: Act as a senior M&A attorney with manufacturing sector experience. Ask: Produce a risk matrix of every seller rep and warranty, rate each high/medium/low, and flag any narrower than ABA model provisions. Format: Table with columns for Section, Rep/Warranty Summary, Risk Rating, Issue, and Recommended Redline. Tone: Analytical and precise, for a partner review memo."

Beyond CRAFT: Six Techniques That Sharpen Any Legal Prompt

CRAFT is the foundation. These techniques separate expert legal prompters from beginners once the five elements are in place.

1. Be Specific About Jurisdiction and Governing Law

Legal analysis is jurisdiction-dependent. "Analyze this non-compete" produces generic output; "analyze this non-compete under California Business & Professions Code Section 16600 and recent Ninth Circuit decisions" produces precise output. The more specific the jurisdictional context, the sharper the analysis.

2. Provide Your Standard or Baseline

When reviewing contracts, do not just ask Claude to "review." Upload your firm's standard terms alongside the opposing party's draft and ask for a gap analysis. A concrete comparison point gives Claude a specific benchmark instead of leaving it to guess at what "market standard" means.

3. Use Chain-of-Thought for Complex Analysis

For multi-step legal analysis, ask for the analysis itself, not just a verdict: "Identify the applicable legal standard, apply each element to the facts, address potential counterarguments, and then state your conclusion." That is far more useful than "Is this claim viable?" Note that on the newest Claude models, such as Claude Opus 5.5, thinking is always on and the model calibrates how much to think to the task, and Anthropic's prompting guidance says a general instruction such as "think thoroughly" often beats a hand-written step-by-step plan, so use the structure to shape the written output rather than to script Claude's reasoning.

4. Specify What You Do Not Want

Ruling things out helps, but Anthropic's own prompting best practices recommend telling Claude what to do instead of only what not to do. Pair each exclusion with the behavior you want: "Do not provide a general overview of contract law. Go directly to the specific provisions at issue." If you do not want hedging: "Provide a definitive assessment rather than listing arguments on both sides." A constraint that names the replacement behavior sharpens output.

5. Iterate, Do Not Start Over

Your first prompt rarely produces the final product. Use the conversation to refine: "That analysis is good but too general on damages. Expand the damages analysis to separately address compensatory, consequential, and punitive damages under Texas law, with statutory caps where applicable." In our experience, iterating gets further than hunting for one perfect prompt.

6. Verify Before It Leaves Your Desk

Prompting is only half the job. Every citation and legal conclusion Claude produces gets independently verified before it reaches a client or court. A perfectly crafted prompt does not change that obligation.

Common Mistakes to Avoid

• Skipping Context: Without jurisdiction and facts, Claude defaults to generic reasoning.
• Vague Ask: "Help me with this contract" is a wish, not a prompt.
• Too broad: "Tell me everything about IP law" produces a textbook, not usable work product.
• No Format: Claude picks its own structure if you don't, and it may not fit how you will use the output.
• No Tone or audience: A motion brief and a client email need completely different registers.
• Forgetting to verify: The best prompt in the world does not make citations self-verifying.