Claude for Legal: A Practical Guide to the New Artifacts Gallery

Anthropic’s new legal artifacts gallery offers a different starting point for legal work: an editable interface for a recurring task, such as comparing contract terms, checking citations or tracking matters. A lawyer can adapt the workflow around the documents, rules and people involved. A business owner can use a suitable workflow to organize questions for counsel. Neither should treat a polished table as proof that its legal conclusions are correct.

Anthropic’s October 28, 2026 announcement adds a gallery to Claude for Legal, launched in May; it reports over 60 legal connectors added since then. Confirm provider subscriptions separately.

Our assessment: the most useful change is the chance to turn a good review method into a repeatable interface. The hard work remains defining the task, checking the evidence, protecting confidential information and deciding who has authority to act. The right first project is a narrow workflow whose output a qualified person can check.

How we prepared this guide: We reviewed Anthropic’s announcement, current product and sharing documentation, partner material and primary legal-ethics and research sources. We did not test the gallery templates on client matters or measure their accuracy.

What Claude’s legal artifacts are

An artifact is a reusable output made in Claude, such as a document or interactive tool. For legal work, think of an intake screen, comparison table or review board that can be adapted for a defined job. Anthropic describes its examples as editable and says their outputs link to sources. Those links are valuable starting points for review, but we have not independently tested whether they reliably support each conclusion. See the legal teams page and the artifact documentation.

Four layers matter. Claude provides the AI capability; the artifact presents a workflow; source documents and connected systems provide information; people decide what the result means and what action to take. A review interface can faithfully quote a clause and still miss an exception in an appendix. A real citation can support a different proposition from the one the output claims.

A connector gives an authorized workflow a route to another service or its information. Anthropic’s legal page lists iManage, NetDocuments, Ironclad and Bloomberg Law as examples. A plugin, as Anthropic describes it, combines tools, skills and integrations for a use case. An API lets developers integrate services into software under separate usage and access terms. Word-based drafting tools are another way to work in documents; a legal artifact is a separate interface and should not be assumed to have the same editing, permissions or subscription features. These pieces can complement one another, but their access rules and costs are not interchangeable.

For legal research, the source collection matters as much as the interface. Lexis+ with Protégé describes research grounded in its legal content and Shepard’s signals. Thomson Reuters describes research using Westlaw and Practical Law materials. Those are vendor descriptions, not an independent ranking. Ask what sources, jurisdictional coverage, updating, citator and reviewer a particular matter requires before choosing tools.

Which legal artifact should you start with?

Anthropic’s legal gallery is the starting point. The following are named examples from the legal solutions page and the launch’s Cite check, not an exhaustive inventory. The review questions are our recommendations, not claims about built-in safeguards.

The launch names Goodwin’s AI Governance Benchmark, Foley & Lardner’s Obligations Checker and PwC’s Subpoena Response Tool: governance gaps, agreement duties and draft subpoena responses, respectively. Harvey Intake is a distinct connected workflow: Harvey says it uses Legal’s materials, routes judgment calls to lawyers in Harvey and returns replies to Claude. Confirm required partner accounts, agreements and pricing.

Anthropic’s announcement names five of nine access-to-justice artifacts: Zealous’s Criminal Charge Reality Check (New York sentencing); Mobile Pathways’ Immigration Papers Reader (dates and missing documents); Rising for Justice’s Rent Ledger Reviewer and Record-Sealing Screener (D.C. ledger errors and sealing eligibility); The Access Project’s Record Clearance (state-specific clearance). Legora and Orrick examples are forthcoming; the Legal Ambassadors Program has a waitlist. Check each tool’s scope and referral route. No measured access-to-justice improvement was established in our review.

Choosing between a table and a playbook board

Use a comparison table when the immediate question is what several agreements say about the same set of topics. A procurement team comparing ten software renewals might need the notice date, price adjustment language and termination rights for each. The output should point to the relevant agreement and provision, and distinguish a clause that was not located from one that affirmatively says no notice is required. It should also name documents the reviewer expected but did not receive.

A playbook board is more useful when the question is whether a proposed provision differs from an approved negotiating position. Suppose the company’s counsel has approved a standard liability cap, a narrower fallback and an exception that requires a senior lawyer’s approval. Supply that version of the playbook and the draft agreement, then ask the reviewer to verify both the clause and the classification. A proposed clause may resemble the fallback yet contain a different carve-out. The board is a starting queue; it cannot grant authority to accept an exception merely by labeling it standard.

Citation checks and jurisdiction surveys

For a citation check, separate three jobs. First, compare the quotation and pinpoint against the original source. Second, read enough surrounding text to decide whether the source supports the proposition in the draft. Third, use an appropriate current-law research process to check subsequent treatment, amendments and controlling authority. A tool may help organize the first two checks, but a match by itself does not complete the third. For example, a correctly quoted decision may have been reversed or may concern a different legal test.

For a jurisdiction matrix, ask one precise question rather than seeking a general survey of “employment law.” State the business activity, jurisdictions, date and assumptions: for example, whether a notice requirement applies to a named category of employer after a specified event. Ask for conditional answers and the facts that could change them. A cell saying “yes, if the employer has at least [threshold] employees” is more useful than an unqualified yes. Local counsel can check disputed interpretation and recent amendments before the matrix informs a decision.

Matter boards and intake

A matter board can make ownership and next steps visible across a team. It is not the authoritative docket. When a hypothetical filing date appears on a board, the assigned reviewer should confirm the triggering event, applicable rule, extensions and calendar entry in the system the organization uses for deadlines. Preserve the underlying order or notice and the calculation, including who checked it. A copied date can be accurate in the original matter and wrong in a later one.

For an intake workflow, distinguish an answer drawn from approved internal guidance from advice about a new legal issue. A routine question about where to find the company’s signed NDA process may be appropriate for a documented answer. A request about a threatened claim, regulator inquiry or unfamiliar liability should go to a named person. Test whether an urgent request reaches that person, whether the requester sees the next step and what happens when nobody is available. A form that collects information without a working handoff can hide a problem in a queue.

Testing access-to-justice use

For an under-resourced user, a useful interface needs more than a clear answer. An organization can test whether people can read and navigate it with assistive technology, understand the language, identify the relevant jurisdiction, and reach a human referral when the facts are outside its scope. Test with synthetic cases that include missing paperwork and urgent deadlines; check whether the tool distinguishes uncertainty from ineligibility. Provide a route to correct an error and avoid requesting sensitive information the organization does not need. These are proposed evaluation criteria, not measured outcomes for the announced tools.

Access, setup and price

Start with the official gallery and an example relevant to your task. We could verify the public links but did not inspect the authenticated gallery or run the templates. Individual template eligibility and the exact screens available to your account therefore need confirmation in Claude.

Anthropic’s general artifact feature matrix says creation is available on Free, Pro, Max, Team and Enterprise. App connections and persistent storage in artifacts require Pro, Max, Team or Enterprise, and web or desktop; code and file execution must be enabled. General artifact access does not establish that every legal example, partner integration or source is available on every plan. A connector in ordinary Claude chat and a connection inside an artifact are separate entitlement questions.

Anthropic’s US pricing page, observed October 28, lists the following headline prices. Annual Team and Enterprise figures are per seat per month, billed annually.

Team requires at least two members. The Enterprise help page distinguishes the 20-seat self-serve minimum from a 50-seat sales-assisted minimum. Taxes, limits and terms apply and prices can change. We did not identify a separate legal-template price; that does not establish that a complete connected legal workflow is free.

The administrator guide says artifacts use existing limits for Team and seat-based Enterprise; usage-based Enterprise consumption is billed at standard API rates. Include partner licensing, implementation, review and ongoing maintenance in your budget, not only the Claude seat price.

A sensible first setup

1. Choose one repetitive task and name its reviewer. Write down the jurisdiction, date, intended users, permitted source set and decisions the workflow must never make by itself.
2. Begin with synthetic or appropriately redacted documents. Include a missing fact and a conflicting source so you can see whether the output exposes uncertainty.
3. Adapt fields and instructions to the actual task. Require an explicit “needs review” state when evidence is missing, ambiguous or unreadable.
4. Only connect approved systems after checking the account’s permissions and requested tools. Record which sources were available and which were not.
5. Open the underlying passages, test the sharing audience with harmless material, and have the responsible person approve the workflow before real matters enter it.

These are operating recommendations, not a claim that the gallery has a built-in testing or approval system. If a control is unavailable in the artifact, put it in the surrounding process.

Example: reviewing a vendor contract

Suppose a small company receives a services agreement with an order form, data-processing addendum and later amendment. Its first objective is not a green “safe to sign” badge. It is an inventory of obligations and questions a reviewer can trace to the relevant language.

Start with tabular review and a synthetic agreement. Define the fields: parties, governing law, renewal notice, termination rights, fees, liability cap, indemnities, data handling, assignment and provisions that survive termination. Add the actual documents and a lawyer-approved playbook where permitted. Identify the version and business owner. If the agreement incorporates an online policy or missing schedule, record that gap instead of treating the main PDF as the whole deal.

The following is an original suggested prompt, not a product-supplied instruction or a tested result:

Adapt this for our procurement review. Use one row per agreement. Add governing law, renewal notice, termination rights, liability cap and assignment. Quote the supporting language and identify the document and page where available. Mark missing, ambiguous or unreadable terms “needs review.” Do not infer an answer from market practice. List missing schedules and conflicts between documents.

The table below illustrates the kind of review record to aim for. The wording and circumstances are invented for this guide; they are not output from a Claude test, actual contract terms or legal advice.

Each row should preserve a route back to the actual source and show the limits of the finding. If an extraction omits the amendment, a reviewer comparing the source pack against the table can catch the discrepancy before the date reaches a calendar. If the system cannot display a field, keep the evidence record in the surrounding matter process.

Now inspect the output against the source. A table may correctly identify a 12-month liability cap while overlooking an indemnity carve-out elsewhere. Defined terms, exceptions, order-of-precedence clauses and amendments can change the result. A reviewer should check the entire relevant provision and related sections, not just the highlighted sentence. A missing page reference calls for manual investigation, not a guess.

Route material deviations to the lawyer and business decision maker authorized to accept them. Record whether each issue was accepted, revised, rejected or sent for specialist advice, together with the relevant document version. Do not send a negotiation position or sign solely because an interface marks a row complete.

A business owner can turn the result into a focused conversation: “Which claims fall outside the cap? What notice must we give before renewal? Which obligations continue after termination?” An unflagged clause is not proof that a contract is enforceable or appropriate for the business.

Example: a dated regulatory scan

A regulatory scan needs a narrower brief than “tell me all the laws that affect us.” Specify the activity, jurisdictions, regulator list, time window and official sources. For each candidate change, record the source, publication date, legal status, effective date, enforcement date, covered entities, transition rules and the assumptions used to assess applicability.

Limit this scan to [business activities] in [jurisdictions] as of [date]. Separate proposals from enacted measures. Show publication, effective and enforcement dates separately. Link the original authority, state the applicability assumption and flag matters requiring counsel’s interpretation. List inaccessible sources and unresolved conflicts.

This is an illustrative prompt, not a test of the gallery. A reviewer must open the authority, check amendments and exemptions, and decide whether it applies. A proposal accurately summarized as though already effective can lead to the wrong operational decision. A “no changes found” report after a source failed to load is also incomplete. Both cases need an explicit escalation.

Assign an owner to each confirmed action and keep the scan date visible. A reusable interface does not itself prove that a scan runs on a schedule, covers every source or sends alerts. Verify any separate monitoring and notification mechanism before relying on it for deadlines.

Sharing, confidentiality and privilege

Sharing a workflow is not the same as giving every viewer access to the same data. Current Claude sharing documentation says artifacts start private. Team and Enterprise organization sharing, owner-controlled external sharing and named outside invitations have different rules. New shared artifacts require a Claude account; published legacy chat artifacts have an anonymous-viewing exception. Named outside email invitations are in beta and have a separate organization setting. Claude-powered or app-connected artifacts cannot use “anyone with the link.”

According to that documentation, outside invitees cannot operate an artifact’s Claude or connector functions. Connected data uses the viewer’s own access, so a source may be unavailable to another person. Shared stored data can nevertheless be visible to everyone opening the artifact. Some legacy Team or Enterprise chat artifacts can share chat attachments as well. Test the audience and contents with harmless material before using real documents; check the current settings because these distinctions can change.

Each user connects their own apps, and tool permission may be requested on first use. Anthropic says tools requiring approval on every action are unavailable within artifacts. Administrators control organization settings, but the admin guide says connector permission is organization-wide rather than a control that restricts each reachable artifact app individually.

Before entering confidential information, review your actual plan and agreement, retention and deletion, training settings, administrator access, source-system permissions, residency, subprocessors, sharing and any client instructions. Anthropic’s legal-work guidance describes commercial customer content as not used for training by default, while discussing configuration and limited safety review. It is the vendor’s perspective, with its discussion of legal authorities stated as of June 2026. Do not extend that commercial default to every personal-plan configuration. It does not guarantee privilege or substitute for matter-specific advice. API zero-data-retention options do not establish the gallery’s retention terms. Separately, Anthropic’s artifact admin guide says artifacts are unavailable through third-party cloud platforms.

Confidentiality, training, retention, privilege and accuracy are separate questions. A favorable answer on one does not settle the others. A lawyer should assess the specific information, client agreement, provider terms and applicable law before use.

What this means for lawyers and small businesses

In-house legal teams

Contract intake and playbook review are natural candidates because requests repeat and business teams need to know who owns the next step. A useful intake captures the counterparty, jurisdiction, deadline, requested decision and source documents, then routes exceptions to the right person. It also needs a fallback for urgent requests, missing policies and questions outside its scope.

Use only the attached approved guidance for routine answers. Collect the requesting team, jurisdiction, counterparty, deadline and source documents. Escalate litigation, regulatory contact, unusual liability and missing policy coverage to [role]. Show the owner and whether the request is waiting on the requester. Do not promise a legal outcome.

This proposed prompt requires testing and operational controls. The legal team still needs to decide who can approve risk, keep guidance current and reconcile the queue with its existing systems. A generated answer is not approval to enter a contract.

Small firms and solo lawyers

A firm can pilot one recurring matter type: extract terms from a known class of agreements, prepare a document checklist or organize a first-pass issue list. Standardization can make omissions easier to spot, but only if someone owns the playbook, checks its updates and reviews exceptions. Solos should account for the fact that the same person may build, operate and verify the workflow. A second reviewer or periodic independent sample check may be useful for consequential work.

The relevant economics are the full cost of an accurate result. If a first draft takes five minutes but requires an hour of correction, the apparent speed is misleading. Setup, source cleanup, review, subscription, partner charges, client communication and change management all count.

Small-business owners

Use an artifact to organize facts, find questions and prepare for a conversation with qualified counsel. It may help compare renewal language across supplier agreements or make a list of documents needed for a lawyer’s review. Check that your account and permissions permit the intended use before uploading sensitive information.

For a first review, gather the signed agreement or proposed draft, all order forms and schedules, amendments, referenced policies, correspondence changing commercial terms, and any renewal notices. Identify the current version and ask which attachments are missing. Start with a synthetic example or an appropriately redacted copy after checking whether redaction is sufficient for the information involved. Keep an untouched original so a lawyer can inspect context that a summary may omit.

Then make a short issue list in business language: when can the price change, how and when can either side terminate, what happens to customer data, what happens if service fails, and which liabilities are uncapped? Note the dates and the person who needs to decide each question. Ask counsel to verify consequential provisions and jurisdiction-specific issues, especially if the business faces an imminent deadline, a dispute, unusual indemnity, a personal guarantee or a regulatory obligation. Keep the signed version and final advice with the business’s normal records. The review should lead to a human decision, not automatic signature.

Be careful with outputs that sound definitive. An AI summary may miss a local rule, an exception or an unfavorable clause altogether. If a contract, dispute, regulatory issue or deadline could materially affect the business, ask a lawyer who knows the relevant jurisdiction and facts to review it.

Jobs, fees and the division of work

Reusable interfaces may shift some first-pass extraction and intake toward standardized processes, leaving lawyers more time for exceptions, negotiation and judgment. They may also create work maintaining instructions, checking sources, managing access and supervising outputs. Which effect dominates depends on the matter and organization. The available evidence does not establish a forecast of job losses or a guaranteed reduction in legal bills.

There is also a training question. Junior lawyers often learn by extracting clauses, researching authorities and comparing drafts before a senior lawyer reviews their work. If a tool performs part of that first pass, a firm may need to teach those skills deliberately: have trainees make an independent assessment, compare it with the generated output, identify both missed and invented issues, and explain their reasoning. This is an editorial recommendation, not evidence that a particular training model has succeeded.

For an in-house team, a better intake process could make service more predictable if requesters see ownership and escalation. It could also shift administrative work to legal operations or create a second queue if it is disconnected from the matter system. Firms may face pressure to explain why a repeatable task costs what it does and to offer different pricing or more proactive services. Whether a client saves money depends on the fee agreement, volume, quality controls and the cost of work still performed. These are possible changes in incentives, not a forecast of adoption or fees.

Clients should ask what work is automated, who reviews it, how errors are corrected and how fees are calculated. Under the ABA’s Formal Opinion 512, lawyers must charge reasonable fees; hourly charges reflect actual time, including appropriate review. A flat fee remains subject to reasonableness. Applicable rules and engagement terms vary.

What the evidence says about legal AI accuracy

Legal AI can produce confident errors even when it uses retrieved sources. A 2024 evaluation of then-current legal research products reported hallucinations in its tested tasks, despite improvements relative to a general chatbot. Its 17–33% hallucination range across the tested queries is tied to those products, versions, questions and definitions; it is not a score for this Claude gallery or a current comparison of today’s legal platforms.

A randomized study published in 2024 found law students using GPT-4 completed selected tasks faster, with small and inconsistent quality improvements. A separate 2026 publication reported an October 2024 experiment using o1-preview and a 2024 version of Vincent AI. Of 137 law students completing at least one of six tasks, 125 completed all six. Aggregate quality and several productivity measures improved, but the NDA task did not show a demonstrated productivity benefit. These are useful reasons to test a workflow, not estimates of performance on your matters or proof of practicing-lawyer outcomes.

Checking a source means more than confirming that a link opens. Does it support the proposition? Is it the right jurisdiction and date? Has it been amended, overruled or qualified? Were contrary authorities and missing documents considered? A correct quotation can coexist with an incorrect legal conclusion.

Professional duties and court rules

ABA Formal Opinion 512 discusses competence, confidentiality, supervision, communication and fees under the Model Rules. It calls for appropriate verification and addresses informed consent for specified risks involving client information; it does not impose a universal consent rule for every AI use. Lawyers should consult the rules, court orders, client instructions and engagement terms that actually govern their work. The New York City Bar’s opinion offers jurisdiction-specific discussion of related duties.

Disclosure to a court is a separate issue from disclosure to a client or consent to a data practice. Canada’s Federal Court notice requires a declaration for covered litigation materials containing directly AI-generated content, with defined exclusions. It is a rule for that court, not every Canadian court. Check the applicable court’s current requirements and verify all authorities before filing.

How to run a pilot that produces useful evidence

1. Choose a bounded sample. Pick a recurring task with a human-reviewed baseline. Include typical documents and difficult cases: an appendix, amendment, negation, conflicting dates and a deliberately missing source.
2. Define success before running it. Record correct material issues, missed issues, unsupported claims and false alarms. Measure setup, source preparation, review, corrections and total turnaround separately.
3. Test boundaries. Include an irrelevant jurisdiction, ambiguous date, unreadable page and misleading instruction embedded in a document. The desired response is a visible uncertainty or escalation, not a confident answer to every field.
4. Trace the evidence. Open the cited passages and inspect exceptions, definitions, tracked changes, tables, footnotes and incorporated references. Log inaccessible sources.
5. Check permissions. Test who can view, modify, copy or share the artifact and stored content. Review source-system access with an administrator before a real matter.
6. Assign accountability. Name the reviewer, decision owner and escalation route. Preserve the output, source set, date and workflow version. Retest after a material change.

There is no universal safe accuracy percentage. One missed deadline or uncapped obligation may matter more than dozens of correct low-risk rows. Compare the complete cost and quality of the workflow with the existing process before expanding it.

Frequently asked questions

Is Claude for Legal free?

General artifact creation is documented on Claude’s Free plan. App connections and persistent storage inside artifacts require an eligible paid plan, and specific gallery or partner entitlements were not verified. Partner products and usage may carry additional costs. Check the current plan and provider terms.

Do lawyers need to code?

Anthropic presents editable examples that can be adapted through Claude. A narrow starting workflow can be specified in plain language, but more complex integrations and organizational controls may need technical help. We did not test the templates ourselves.

Can I send a client a link?

Sharing depends on the artifact type, account, organization settings and invitation method. See the sharing details above and check stored data and access before inviting anyone.

Does an artifact give legal advice?

It can generate legal-sounding explanations, but its answer is not a substitute for advice from a qualified lawyer who can evaluate the facts and applicable law. A business owner should treat it as a preparation aid and seek advice on material decisions.

Can I rely on its citations?

Use citations as leads. Open the original source, verify the quoted passage and proposition, and check current status and jurisdiction. No independent accuracy rate for these gallery templates was established in our review.

Do the access-to-justice tools work everywhere?

Check the jurisdiction limits listed above and the individual tool’s current scope. Find appropriate local assistance when necessary.

Will a regulatory artifact keep me compliant automatically?

A scan can organize potential developments for review. It does not establish complete source coverage, continuous execution, applicability or a completed compliance action. Confirm the source, date, legal status and owner for each item.

Right now

Trending on Kingy

Keep reading with the stories getting the most attention now.

1. 01 Blog · Oct 20

   Claude Fable 5.5 Live: Launch Updates, Specs, Benchmarks & Comparisons

   Read story →
2. 02 Blog · Oct 26

   GPT-6.1 Terra: Benchmarks, Pricing, Specs, and How It Compares

   Read story →
3. 03 Blog · Oct 23

   Dessa AI: Jack Dorsey’s Block Built a Voice Model Without Text

   Read story →

The Kingy Brief

Get The Kingy Brief.

AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.

Free · Double opt-in · Unsubscribe anytime

Privacy policy

Sign up for The Kingy Brief on Beehiiv.

Signup help and newsletter schedule

Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.

Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.

Sign up free