Self-Audit Your Claude Use Against the Ethics Rules and Draft an AI-Use Policy
Inventory how you and your team actually use Claude, map each use to the core ethics rules and the text of ABA Formal Opinion 512, rank the gaps, and draft a firm AI-use policy and client disclosure language.
When to use this
Use this when you have been using Claude for a while and want to check your practices against your professional obligations, or when a client, insurer, or colleague asks what your AI policy is. It works for a solo reviewing their own habits and for a legal team writing its first written policy. Claude organizes the analysis from the materials you paste; the lawyer confirms it against current rules and [JURISDICTION] guidance.
Most lawyers adopt Claude one task at a time: a summary here, a first draft there, a research question late at night. Months later, nobody has written down which plan the work runs on, what client data goes in, who checks the output, or how the time is billed. Those are exactly the questions an ethics inquiry, a client questionnaire, or a malpractice carrier will ask.
This workflow turns that drift into a record. You list how Claude is actually used, then have Claude map each use to the duties ABA Formal Opinion 512 discusses for generative AI: competence (Model Rule 1.1), client communication (1.4), fees (1.5), confidentiality (1.6), meritorious claims and candor to the tribunal (3.1 and 3.3), and supervision of lawyers and nonlawyers (5.1 and 5.3). You paste the opinion in full so Claude quotes it rather than paraphrasing from memory. From the gaps it finds, you draft a firm AI-use policy and client disclosure language.
Claude is organizing your analysis, not issuing an ethics opinion. State rules and ethics opinions differ, and they change, so every state-specific point stays as a [JURISDICTION] placeholder until you research it. Court disclosure requirements are flagged separately because they can come from local rules and individual judges' standing orders. The lawyer confirms every conclusion before anything is adopted.
The Workflow
Inventory how Claude is used
Write down every recurring use, honestly. For each, note the task, what data goes in, which Claude plan it runs on, who reviews the output, and how the time is billed. If you have a team, ask each person for their list. Then have Claude organize it and ask about gaps.
PromptI am reviewing how [I / our legal team] use Claude, as part of an ethics self-audit. We are [DESCRIPTION, e.g., a litigation practice / an in-house legal team at a manufacturing company] practicing in [JURISDICTION]. Our Claude plan(s): [PLAN NAMES AND WHO USES EACH]. <document name="usage_inventory"> [PASTE YOUR LIST: task | data that goes in | plan | who uses it | who reviews output | how it is billed] </document> Organize this into a table with those columns. Then list any use where a column is blank or vague, and ask me up to 8 questions to complete it, for example whether client names are included, whether citations are checked against the source, or whether a nonlawyer uses it without review. Do not analyze the ethics rules yet.
What you get: A complete usage table and a short list of questions. The inventory is only useful if it reflects what people actually do, so answer the questions candidly.
Map each use to the rules, quoting the text you paste
Paste the text of the Model Rules you are using and ABA Formal Opinion 512, plus any [JURISDICTION] rules or opinions you have already found. Ask Claude to map each use to the duties and to quote the pasted text for every point. Anything it cannot support from the pasted text should be marked as an open question.
PromptBelow are the rule texts and ethics guidance I want to audit against. <document name="model_rules"> [PASTE TEXT OF MODEL RULES 1.1 (WITH COMMENT 8), 1.4, 1.5, 1.6, 3.1, 3.3, 5.1, AND 5.3, OR YOUR JURISDICTION'S VERSIONS] </document> <document name="aba_formal_opinion_512"> [PASTE THE FULL TEXT OF ABA FORMAL OPINION 512] </document> <document name="jurisdiction_guidance"> [PASTE ANY STATE RULES OR ETHICS OPINIONS ON GENERATIVE AI, OR "none yet"] </document> For each use in the inventory, return: Use | Duty implicated (competence, communication, fees, confidentiality, candor and meritorious claims, supervision, other) | Relevant text (exact quote with rule or section reference from the pasted documents) | How our current practice measures up | Open questions. Quote only from the pasted documents. Do not paraphrase rules from memory. Where the pasted documents do not address a point, write "Not addressed in pasted text" and, for state-specific points, "[JURISDICTION]: research needed".
What you get: A rule-by-use map with quotes you can check. Any "Not addressed" or [JURISDICTION] entry is a research item.
Rank the gaps
Have Claude turn the map into a ranked list of practices that need to change. Typical gaps: client data on a plan without commercial terms, citations that are not verified against the source, nonlawyer use without review, and billing practices that do not reflect actual time spent.
PromptFrom the map, list every practice that may fall short of the quoted text. For each give: Gap | Uses affected | Quoted text it relates to | Risk ranking (High, Medium, Low) with a one-sentence reason | Suggested change | Owner [TO ASSIGN]. Rank High where client confidential information or a court filing is involved, or where output reaches a client or court without lawyer review. Sort by ranking. Do not soften a gap because it would be inconvenient to fix, and do not overstate one the quoted text does not support.
What you get: A short, sorted gap list with concrete changes. Agree the owners and dates with your team before drafting the policy.
Draft the firm AI-use policy
Use the gap list to draft a written policy. Keep it short enough that people will read it and specific enough to follow: which tools and plans are approved for what, what data may go in, how citations are verified, who supervises, how use is billed, and what training is required.
PromptDraft an AI-use policy for [FIRM OR LEGAL TEAM NAME] that addresses every gap on the list. Sections: 1. Purpose and scope (who and what it covers) 2. Approved tools and plans, and what each may be used for [FILL FROM OUR INVENTORY] 3. Data handling: what may and may not be entered, by plan 4. Verification: every citation, quotation, and factual statement checked against the original source before use; no unverified authority in any filing or client advice 5. Supervision: lawyer review of all output; responsibilities when nonlawyers use the tools 6. Billing: [OUR APPROACH, TO BE DECIDED BY THE BILLING PARTNER], with reference to Model Rule 1.5 and the fee discussion in ABA Formal Opinion 512 as quoted in our map 7. Client communication and court disclosure (see separate section) 8. Training and review date Plain language, numbered sections, under 2 pages. Use [BRACKETED PLACEHOLDERS] for decisions we have not made. After the draft, list each placeholder.
What you get: A two-page draft policy with open decisions marked. The responsible partner or general counsel decides each placeholder.
Draft client disclosure language and flag court rules
Draft engagement-letter language and a short client FAQ explaining how AI is used on client matters. Court disclosure is separate: some courts and judges require disclosure or certification of AI use, so those obligations stay flagged until checked for each court.
PromptDraft two client-facing pieces consistent with our AI-use policy: 1. An engagement-letter paragraph (under 150 words) explaining that we may use generative AI tools on the matter, how client information is protected, that a lawyer reviews all work, and how use affects billing [PER POLICY SECTION 6]. Do not promise anything the policy does not require. 2. A client FAQ of 5 to 7 questions with short answers (what we use, what happens to your information, who checks the work, how it affects your bill, how to raise concerns). Then add a separate checklist titled "Court disclosure obligations" with one line per court where we file: [LIST COURTS], each marked [CHECK LOCAL RULE AND STANDING ORDERS]. Do not state any court's requirement.
What you get: Draft engagement language, a client FAQ, and a court-by-court checklist to research before the next filing in each court.
Lawyer review and follow-up
Confirm the analysis against the current text of the rules and against [JURISDICTION] ethics opinions, resolve each policy placeholder, and set a review date. Record the topics this raised (for example, data handling by plan or citation verification) as candidates for CLE or internal training. The policy and disclosures are drafts until a responsible lawyer adopts them.
Example Output
Illustrative example — names, facts, and figures are fictional.
GAP LIST (excerpt; fictional practice: Ostrander Legal, two lawyers and one paralegal) 1. HIGH | Client documents summarized on a consumer plan with training not turned off Uses: deposition summaries, medical-record review Relates to: Model Rule 1.6(c) as pasted; Op. 512 confidentiality discussion as pasted [quote in map, row 3] Change: move client work to a plan whose commercial terms exclude training, or confirm privacy settings in writing; no client documents on personal accounts. Owner: [TO ASSIGN] 2. HIGH | Research memos sent to client without checking citations against the source Uses: research first drafts Relates to: Rule 1.1 and Comment 8 as pasted; Op. 512 competence discussion [map, row 5] Change: every citation pulled and read before any memo leaves the office; log the check in the file. Owner: [TO ASSIGN] 3. MEDIUM | Paralegal uses Claude for discovery responses with no recorded lawyer review Relates to: Rule 5.3 as pasted [map, row 7] Change: lawyer sign-off required before any AI-assisted draft is served. 4. MEDIUM | Flat 1.0 hour billed for summaries that now take 0.3 Relates to: Rule 1.5(a) as pasted; Op. 512 fee discussion [map, row 9] Change: billing partner to set approach; see policy s. 6 placeholder. 5. LOW | No written record of which plan each person uses Change: add to policy s. 2 and review quarterly. OPEN RESEARCH - [JURISDICTION]: any state bar opinion on generative AI and client consent - Court disclosure: [COURT A], [COURT B] marked [CHECK LOCAL RULE AND STANDING ORDERS] DRAFT FOR ATTORNEY REVIEW. Quotes are from the texts pasted on [DATE]; confirm against current versions.
Tips
- •Paste the actual text of Opinion 512 and the rules. A map built on quoted text can be checked line by line; one built on Claude's memory cannot.
- •Be candid in the inventory. The audit is only as useful as the description of what people really do, including shortcuts taken under deadline.
- •Check which plan each person uses. Consumer plans and commercial plans have different data terms, and that difference drives most confidentiality gaps.
- •Keep court disclosure separate from client disclosure. Client communication is governed by the ethics rules; court disclosure can come from local rules and individual judges' orders.
- •Write the policy for the people who will follow it. Two clear pages beat ten pages nobody reads.
- •Re-run the inventory step on a set schedule. New uses creep in, and the policy should follow them.
A note on confidentiality
The usage inventory should describe kinds of data and tasks, not include actual client documents or names. Do the audit on the plan you intend to approve for client work, and record which plan that is. The rule map, policy, and disclosure language are drafts for a responsible lawyer to adopt: every quotation must be checked against the current official text of the rules and ABA Formal Opinion 512, and every [JURISDICTION] and court item researched before the policy takes effect.
Frequently Asked Questions
Why paste ABA Formal Opinion 512 instead of asking Claude what it says?
Because the audit has to rest on text you can check. When Claude quotes the pasted opinion, you can verify each point in seconds. A paraphrase from memory may be incomplete or wrong, and you would have no way to tell without reading the opinion anyway.
Which ethics rules does this workflow cover?
It maps uses to the rules ABA Formal Opinion 512 discusses: competence (Model Rule 1.1), communication (1.4), fees (1.5), confidentiality (1.6), meritorious claims and candor to the tribunal (3.1 and 3.3), and supervision (5.1 and 5.3). Your jurisdiction's versions of these rules and any state ethics opinions control, so add them in Step 2 and keep state-specific points as [JURISDICTION] items until researched.
Do I have to tell clients I use Claude?
It depends on the use, the client, and your jurisdiction. Opinion 512 discusses this under the communication duty (Model Rule 1.4) and confidentiality (1.6). Map your uses against the pasted text, check [JURISDICTION] guidance and any client requirements, and decide with the responsible lawyer. The disclosure language in Step 5 is a draft to adapt once you have decided.
Do courts require disclosure of AI use in filings?
Some courts and individual judges do, through local rules or standing orders, and requirements differ. This workflow lists each court you file in with a [CHECK LOCAL RULE AND STANDING ORDERS] flag so you research each one rather than relying on a general answer.
Prompts used in this workflow
Get New Workflows Like This Every Week
Join the free Claude for Lawyers newsletter — practical workflows, prompts, and guides for everyday practice.