Every senior attorney knows the moment: you've just finished a 90-minute client call, your notes are a mix of shorthand, starred questions, and half-formed strategy threads, and you need a draft engagement letter, a memo, or a set of talking points in your hands before your next meeting. The work is real. The time isn't.
The short version: Using Claude, attorneys can move from raw, lightly anonymized client notes to billable-quality first drafts in a fraction of the time it would take to start from a blank page โ without sending client identifiers to a third-party server, without abandoning professional judgment, and without producing output that reads like it was generated by a machine. The Leverage Years teaches this exact workflow, designed for practicing attorneys who want practical leverage on their actual caseload, not a seminar on AI theory.
The key constraint โ the one that separates a workable attorney AI workflow from a malpractice risk โ is how you handle what you put in. This article walks through exactly that: what to include, what to strip, how to structure the prompt, and how to review the output before it earns a billing code.
If your bar association has issued guidance on AI tools (most now have), this workflow is designed to be compatible with that guidance. It does not substitute for your independent review of applicable ethics rules and your firm's specific policies.
Who this is for
- Practicing attorneys at the associate, senior associate, or partner level who personally touch drafting work โ not just those who supervise it.
- Solo and small-firm practitioners who don't have a first-year associate to hand rough notes to.
- In-house counsel who draft communications, policies, and board summaries under time pressure.
- Legal department leads who want a consistent drafting standard across their team.
This is not for you if: you are looking for AI to replace independent legal analysis, conduct case research, or interpret legal standards without your review. Claude is a drafting accelerant, not a legal reasoning engine.
The Workflow: From Raw Notes to Reviewable First Draft
The workflow has six steps. The actual Claude interaction happens in steps three through five. Steps one, two, and six are yours alone โ and they're where your professional judgment lives.
Step 1: Take the call without restraint
Don't modify how you take notes during the client conversation. Your existing shorthand, margin annotations, and rough strategy threads are exactly right. The goal is to capture the substance of what was said and decided. Fix-up is not the attorney's job in real time.
What you're looking for in your notes afterward: the matter type, the key facts, the client's stated objective, any time-sensitive elements, and the specific document or communication you need to produce.
Step 2: Anonymize before you prompt
This step is non-negotiable. Before any content touches Claude, you strip or generalize all client-identifying information.
The standard substitution set:
- Client name โ "Client A" or a generic descriptor ("the corporate client," "the individual plaintiff")
- Opposing party โ "Party B" or "Counterparty"
- Jurisdiction-specific identifiers (court case numbers, property addresses, entity EINs) โ generic placeholders ("a Delaware LLC," "a commercial property in a northeastern U.S. city")
- Dates that could identify a matter โ relative timing ("approximately eight months ago," "in the prior fiscal year")
- Dollar amounts โ order-of-magnitude placeholders ("a mid-six-figure sum") unless the precise figure is necessary for the drafting task
What you do not need to strip: legal standards, procedural posture descriptions, the nature of the cause of action, the structure of a deal. These are the substance that Claude uses to produce useful output. The identifier is what creates the confidentiality exposure.
If your firm has an enterprise Claude account with a BAA or the equivalent confidentiality commitment, your anonymization requirement may be less stringent than it is when using a standard consumer subscription. Check your firm's current policy. The anonymization practice described here is the conservative baseline that works under any configuration.
Step 3: Build the role + task prompt
The quality of the output depends almost entirely on how clearly you specify the role and the task. Attorneys often under-specify because they're used to delegating to people who already share professional context. Claude doesn't have that context unless you provide it.
A strong role prompt for a drafting task looks like this:
You are a senior associate at a corporate litigation firm. You have strong drafting instincts, understand legal memo conventions, and write in a clear, formal, matter-of-fact style without hedging unnecessarily. You do not fabricate legal citations. You do not make assumptions about facts I have not provided. If something is unclear, you note it rather than inventing an answer.
Then the task prompt:
Using the notes below, draft a client update memo for a matter involving [the anonymized description]. The memo should: (1) summarize the current procedural posture in plain language, (2) explain the three options we discussed with a one-paragraph description of each, (3) identify the next step and its deadline, and (4) close with a professional but warm sign-off appropriate for a long-standing client relationship. Use formal legal memo format. Keep the total length under 600 words. Here are my notes: [paste anonymized notes]
That is a fully specified prompt. The role, the format, the structure, the length, the tone, the constraints โ all declared upfront. The output you get back will be proportionally better.
Step 4: Review the first output structurally
When Claude returns the draft, read it first for structure, not substance. Ask: Does the memo follow the format I requested? Does it have the right number of sections? Is the length in range? If the structural answer is yes, proceed. If not, send a follow-up prompt correcting the specific structural issue before you get into substantive edits.
This saves time because you are not editing a document that has the wrong architecture. Fix the architecture first.
Step 5: Do one substantive review pass
Now read the draft as a professional. You are looking for:
- Legal accuracy: Does anything in this draft assert a legal standard, describe a procedural rule, or characterize a case posture in a way that is incorrect or incomplete?
- Factual fidelity: Does the draft accurately reflect what was in your notes, or has it elaborated in ways that weren't intended?
- Tone and relationship fit: Does this read the way you would speak to this particular client? Claude produces a competent default tone, but you know the client.
- Missing elements: Is there anything your notes captured that didn't make it into the draft?
The review pass is not optional. It is the billable work. Claude is producing a first draft that compresses the blank-page phase; your review is what makes it professional work product.
Step 6: Finalize, re-identify, and bill
Once the substantive review is complete, restore the actual client name, case identifiers, and proper nouns. The document that leaves your desk has real names and real accuracy. Run your standard proofing protocol. Bill appropriately for drafting and review time โ the time you actually spent, which will be shorter than it used to be.
Checklist: Pre-Prompt Anonymization Review
Before pasting any notes into Claude, confirm each item:
- Client name replaced with generic identifier
- Opposing party name replaced or generalized
- Case number or docket reference removed
- Property addresses generalized or removed
- Entity names (companies, trusts, LLCs) replaced with generic descriptors
- Individual names other than client replaced or removed
- Jurisdiction-specific details generalized where not required for the drafting task
- Dollar amounts at appropriate level of abstraction for the drafting purpose
- No Social Security numbers, EINs, or financial account identifiers present
- The result: if this text were intercepted, it would not be attributable to a specific client or matter
This is a two-minute pass. It becomes automatic within a week of practice.
What Claude handles well โ and what it doesn't
Use Claude for:
- Drafting engagement letters, retainer communications, and matter update memos
- Structuring a first draft of a brief's argument section based on your legal outline
- Converting rough client meeting notes into a formal summary or chronology
- Producing multiple versions of a section at different lengths or tones so you can choose
- Drafting FAQs, client explainers, and plain-language summaries of complex legal positions
- Reorganizing a disorganized argument outline into a logical hierarchy
Never use Claude for:
- Primary legal research โ it will hallucinate citations, statutes, and case holdings with complete confidence
- Any work where the actual client-identifying information must be present in the prompt (absent a firm BAA or enterprise agreement)
- Anything you would not personally review before it leaves the office
- Generating legal strategies based on Claude's independent analysis rather than your own
The second list isn't Claude's failure โ it's the correct division of labor between a drafting tool and a practicing attorney.
The compounding effect: building a prompt vault
The attorneys who get the most value from this workflow are not the ones who use it occasionally. They're the ones who document their prompts.
After you get a result you're satisfied with, save the role prompt and the task structure (stripped of any matter-specific content) in a notes file or a simple document. Label it: "Corporate litigation update memo โ standard," "Estate planning engagement letter โ first meeting," "IP dispute timeline summary โ opposing party narrative."
Within a few months of consistent practice, you have a prompt vault: a curated library of starting points for your most common drafting tasks. Each one produces a first draft that is closer to your standard than a generic prompt would. You spend less time in the setup phase and more time in the review phase โ which is where the judgment lives and where the billing properly belongs.
This is what The Leverage Years means by an operating system: not a clever trick you use once, but a structure that makes your standard output faster, more consistent, and more professional over time.
Confidentiality, professional judgment, and your bar obligations
Confidentiality is not an obstacle to using Claude in legal practice โ it is the constraint that shapes how you use it. Treating it as an obstacle leads to shortcuts that create real risk. Treating it as a design constraint leads to a workflow that is both safe and effective.
Several bar associations (New York, California, Florida, and others) have issued ethics opinions or formal guidance on attorney use of generative AI tools. The consistent themes: do not disclose client confidences to AI systems unless you have an appropriate data protection agreement; supervise AI-generated output as you would work product from a first-year associate; do not represent AI output as your independent legal analysis without review.
This workflow is designed around those principles:
- Anonymization before input eliminates the confidentiality exposure at the source.
- Role and task prompting directs Claude toward structured output, not open-ended legal analysis.
- Mandatory substantive review ensures that nothing leaves the attorney's hands without professional judgment applied.
- Prompt vault discipline means that the workflow can be explained and defended if ever required to be.
None of this is theoretical. These are working practices drawn from the specific constraints of legal professional responsibility.
Making this a repeatable operating system
The single-matter workflow described above is useful on day one. What converts it into practice leverage is repeatability:
Standardize your note format. If your post-call notes follow a consistent structure (matter type, key facts, client objective, document needed, deadline), the anonymization pass takes less time and the prompt is easier to build.
Build templates for your most common documents. For any document type you produce more than five times a year, write a reusable task prompt. Store the template; adapt it per matter.
Create a review protocol. The five questions in Step 5 above โ legal accuracy, factual fidelity, tone, relationship fit, missing elements โ should be your standing review pass for any Claude-generated draft. Having a consistent checklist means the review is thorough and fast, not ad hoc.
Log what works. When a prompt produces a draft that requires minimal editing, note what made it work: the role description, the structural instructions, the length constraint. When a prompt produces weak output, note why. Over time, the vault improves.
The attorneys who dismiss this workflow do so because they tried a generic prompt, got generic output, and concluded the tool was inadequate. That is the wrong inference. The tool responds to specificity. Attorneys who are precise in their work are positioned to get more from it than almost any other professional category.
If this is the kind of practice leverage you want to build systematically, The Leveraged Attorney walks through the full workflow, the complete prompt vault organized by document type, the review layer, and how to extend the system to your specific practice area.
Not sure which course fits your work? Take the 6-question course selector.
You can also browse the full course catalog at The Leverage Years courses page.
Frequently asked questions
Does using Claude for drafting create confidentiality or privilege issues?
It can, if client-identifying information is sent to a third-party server without appropriate data protection in place. The workflow in this article addresses that directly: you anonymize before prompting, and you restore real identifiers only after review on your own system. Many firms also use enterprise accounts with confidentiality agreements. Review your bar's specific guidance and your firm's policy โ this article gives you the structural approach, not a substitute for your own professional judgment on applicable obligations.
Will Claude's drafts be detectable as AI-generated?
Drafts produced using the role-and-task prompting approach described here tend to read like competent associate work: structured, clear, formal. They don't have the telltale generic quality of a default ChatGPT output because you've specified the role, format, and constraints. Your substantive review pass also catches places where the tone needs to sound more like you. After review and editing, the finished work product sounds like your office.
How long does the full workflow actually take?
Once the process is familiar, the anonymization pass takes two to three minutes, the prompt setup takes one to two minutes, and the review pass takes five to fifteen minutes depending on the document complexity. You're typically looking at ten to twenty minutes total for a document that would have taken forty-five to ninety minutes to draft from scratch. The compounding benefit comes from the prompt vault โ each reuse of a tested template shortens the setup phase further.
Can this workflow be used for pleadings and court filings?
With additional rigor, yes. The review standard for court filings is higher, because the document will bear your signature and your professional representation of its accuracy. For pleadings, use Claude to draft argument structure and prose, then apply a full legal accuracy review against your own research before anything goes to the court. Never rely on Claude-generated citations or legal holdings without independent verification.
What if my firm has a policy against using AI tools?
Then follow your firm's policy. This article is not an argument that you should circumvent firm governance. If your firm has a blanket restriction, the appropriate path is to understand the reasoning and engage your firm's technology or ethics leadership about how a structured, confidentiality-safe workflow might be evaluated under the firm's criteria. Many firms are actively developing policies; bringing a concrete, documented workflow to that conversation is more productive than asking for a policy change in the abstract.
How does this workflow handle matters where facts are still developing?
Well, actually โ this is a case where Claude's constraints become useful. Because Claude only works with what you give it, a draft produced from notes that reflect a still-developing factual record will be accurate to that record. The document will flag gaps as questions rather than filling them in. Your review pass identifies where the draft correctly notes "pending further information" versus where you need to add known facts that weren't in the notes.