Claude oder Legal AI für Anwälte: Wann Sie bei Claude bleiben und wann HAQQ passt

Start with the honest answer: a lot of legal work does not need us

Claude is a strong model for legal work, and Anthropic is building for lawyers on purpose: its 12 May 2026 release added "20+ new MCP connectors" and "12 new plugins tailored to specific legal work and practice areas", open source and available in Claude Cowork. Every one of those plugins "starts with a short setup interview that learns your practice: your playbook, your escalation chain, your risk calibration, your house style."

Model quality is not a reason to leave either. When we ran three Claude models against ten lawyer-grade prompts in September, we could not pick a winner. And the benchmark table on our Justinian® page describes Justinian®'s result on most rows as "parity with the frontier rather than a lead."

So if you are reading this to find out whether HAQQ writes better sentences than Claude, the honest answer is that the difference sits somewhere else.

Stay on plain Claude if most of these are true:

• Your matters run mainly under US, UK or EU law, which is the law the public plugins mention most.
• You draft in English.
• You already pay for a research tool, you have connected it, and you check every citation before anything leaves your desk.
• You are on a Team, Enterprise or API plan, or you strip client details before you paste. (On Free, Pro and Max, whether your chats train future models is a setting you choose.)
• You like building prompts, skills and practice profiles, and you have the hours to keep them current.

If that is you, stay. This post should not change your mind, and we make the full case for staying in Why lawyers use Claude instead of legal AI, and where it breaks.

The rest of this post is for the lawyer who ticked fewer than three of those boxes, usually because the matter sits in Dubai, Riyadh, Doha or Kuwait City, and the draft has to exist in Arabic.

What we counted in Claude's legal plugins

Everyone writing about Claude for legal work describes the plugins. We wanted to know which law they are written around, so we counted.

Method. On 1 October 2026 we cloned the public repository anthropics/claude-for-legal at its latest commit (4a6c651, dated 23 July 2026). It holds 315 files, 151 of them skill files (SKILL.md), which are the instruction sets the plugins run. We ran a case-insensitive, whole-word search of those files for the jurisdictions they name. Anyone can repeat it in under a minute.

Result. "EU" appears in 28 skill files, "UK" in 16, "Delaware" in 14, "California" in 10 and "New York" in 7.

We do not read this as a flaw. Anthropic's launch post says plainly that "no single set of plugins can cover every practice", and the repository is built to be customised: each plugin is "built around a cold-start interview that learns your playbook and a CLAUDE.md practice profile that every skill reads from." The plugins are a starting point, written around the practice their authors see most.

The research layer follows the same pattern. The repository's own advice is: "Start by connecting a research tool. Everything else is better with one, and citations are unverified without one." The research tools its citation guardrails look for are CourtListener, Trellis, Descrybe and Solve Intelligence, and its connector table describes CourtListener as "Federal dockets and opinions" and Trellis as "State court dockets and motions". The May launch post also lists other research connectors, including Legal Data Hunter and Midpage. A GCC lawyer can wire one in. Choosing it, setting it up and checking what it covers for your jurisdiction stays with you.

What you would build and maintain on your own

If you practise under Gulf law and stay on plain Claude, this is the work you are signing up for. None of it is hard. All of it is ongoing.

A jurisdiction profile

The practice profile is where you tell Claude which law governs, which forum you are in and where the rules split. For a UAE matter that means writing down whether you are before the onshore courts, the DIFC or ADGM, and making the model keep them apart. Our own prompt library puts that rule in one line: "Where onshore and the financial free zones diverge, set them out side by side rather than blending them." The repository says the setup interview takes 10 to 20 minutes per plugin, and that skipping it "is the single most common reason a skill produces generic output." Then the profile has to keep pace with the law, which is the part that is easiest to leave out of the budget.

A research source you trust for your jurisdiction

Citations stay unverified until one is connected. For Gulf law, you pick that source, connect it, learn what it holds and what it does not, and still read the authorities it returns.

An Arabic house style

Claude works in Arabic. Claude's multilingual documentation lists Arabic among the languages it benchmarks and describes "strong cross-lingual performance relative to English." What you build is everything around the language: the legal register your firm drafts in, your bilingual clause pairs, your rule on which language version prevails, and the instruction that makes the model draft in Arabic rather than translate from English.

A confidentiality step

Anthropic's consumer terms cover the Free, Pro and Max plans, where letting your chats train new models is a setting you control. Those terms "do not apply to services under our Commercial Terms, including Claude for Work", which covers the Team and Enterprise plans, or to the API. On any plan, the habit worth keeping is to strip client names before you paste. We have written separately about why AI conversations are not privileged. For the confidentiality question in depth, read Confidentiality Is the Next Hallucination: AI Client Confidentiality for Lawyers.

A team layer

If you work alone, skip this one. With colleagues, someone owns the shared playbooks and decides which plugins everyone runs. Claude supports this at the organisation level: Anthropic's legal deployment guide says plugins, skills and connectors "can be provisioned through admin-managed marketplaces in both Claude.ai and Claude Cowork rather than installed per user." Someone still has to be that admin, and the profiles still have to be written.

The common thread: the model is excellent and the scaffolding is yours. For a lawyer who enjoys that work, it is a feature. For a lawyer who would rather spend those hours on clients, it is the cost.

What HAQQ packages instead

HAQQ packages that same layer and maintains it for you. The model underneath is a third-party frontier model, and our Justinian® page says so: "A third-party frontier model still performs the final generation, and we would rather say so plainly than imply we trained our own." Justinian®, HAQQ's legal AI engine, is model-independent by design. In the Justinian® page's words: "We deliberately do not build on a single vendor."

Here is what that layer covers today, worded as the live site words it.

• Gulf jurisdictions, not a country name on a general answer. Legal AI Chat gives "jurisdiction-aware answers under UAE, Saudi, Egyptian, Qatari, Kuwaiti, Bahraini, Jordanian and Lebanese law." In the site's words: "Tell it you are in the DIFC, in Riyadh or in Cairo and the reasoning, the forum and the citations change with it." The Justinian® page is candid about the limit: "Depth varies by jurisdiction, because it depends on how well the source law is digitised and how much of it is public."
• Arabic as a drafting language. Justinian® drafts in English, Arabic and French, and "Arabic is a drafting language here, not an English answer rendered in Arabic script."
• Retrieval and citations in the same tool. "Justinian® retrieves the statutes, case law and matter documents that actually apply, and cites what it used." You still read what it cites.
• Your drafting style. The Legal AI Twin is the personalisation layer that adapts outputs to your firm's context and style. Load your precedents once and "new drafts come out on your paper, in your format, with your numbering, ready to export to Word".
• Pseudonymisation before the model call. In Legal AI Chat, "personal details in your conversation are swapped for made-up stand-ins before anything is sent to an external AI model provider," and HAQQ puts the real details back before you see the answer. If that step cannot run, the docs say: "Your request is stopped rather than sent with your real details." The docs call it pseudonymisation, not anonymisation, because HAQQ keeps the link between each stand-in and the real detail.
• A workspace built for a team. "Every seat gets its own AI and its own private history," and credits "sit in one workspace pool that everybody draws from." With an eFirm subscription, the AI "also reaches the firm's matters, documents, contacts and billing context."

What you give up

A fair comparison has a cost column too.

• Breadth. Claude is one tool for legal and non-legal work. Our own Claude vs HAQQ comparison says that for writing code, analysing a spreadsheet or researching a market, Claude is "a legitimately strong choice". HAQQ is for legal work.
• Control. On plain Claude you choose the model, write every instruction and fork every skill. In HAQQ, the model choice and the layer around the model are ours to maintain.
• A known cost per request. HAQQ meters usage in credits, and our pricing page says "it is not possible to know the exact credit cost of a request before sending it." A shared pool also means, as our docs put it, "one heavy user can draw most of it down."
• Quality in, quality out. The Justinian® page states: "The engine cannot independently verify factual claims, quality in determines quality out." That holds for HAQQ as much as for Claude.

The decision table

Most readers will land on both sides of this table. Our comparison page says it is "common, and reasonable, to keep both", with HAQQ for the legal work and Claude for the general-purpose jobs.

If you stay on Claude, take these with you

Two free HAQQ Academy resources work alongside Claude, with no account.

• The [prompt library](https://academy.haqq.ai/prompt-library). Several hundred ready-written legal prompts, filterable by jurisdiction, use case, practice area and complexity, each with a Copy button. The UAE liability-cap prompt quoted above is one of them. Paste it into Claude as it is.
• The [redaction tool](https://academy.haqq.ai/tools/legal-redaction-tool). Paste a document, confirm the matches, and replace each name, company, amount or ID with a typed placeholder such as [COMPANY-1] before the text goes into any AI tool. "Detection, redaction and restoration all run locally in your browser." It uses distinct numbered placeholders because a 2026 study (arXiv 2609.11335) found that one generic label hurt model performance more than a distinct pseudonym per entity, though that study tested general reasoning benchmarks, not legal tasks.

HAQQ's take

We run on frontier models too. After our September benchmark we wrote that we "moved routine legal drafting to Opus 5.5", because it was indistinguishable on quality from the two dearer Claude models we tested. We are not selling an alternative to Claude-class models. We are selling the layer around one.

That layer is the whole argument. Generic AI starts with the prompt. HAQQ starts with the organisation, jurisdiction, matter, sources and workflow around the prompt. If you would rather build that layer yourself, Claude gives you excellent parts and an open repository to start from. If you would rather not maintain a Gulf jurisdiction profile on top of your caseload, that maintenance is the work we do.

So the useful question is who maintains everything around the model, and whether that is how you want to spend your week.

Key Takeaways

• Plain Claude is enough for English-language work under US, UK or EU law when you verify every citation yourself. If that is your practice, stay on it.
• The public claude-for-legal skills are written around US, UK and EU practice, so the Gulf layer is yours to add.
• On plain Claude, a Gulf practice builds and maintains five things: a jurisdiction profile, a research source, an Arabic house style, a confidentiality step and a team layer.
• HAQQ packages that layer around a third-party frontier model: jurisdiction-aware answers under UAE, Saudi, Qatari, Kuwaiti and Bahraini law, Arabic drafting, citations, pseudonymisation before the model call and a shared workspace.
• The trade-off is breadth and control on Claude against maintenance you do not do on HAQQ. Keeping both is common.

Sources & further reading

• Claude blog: Claude for the legal industry (12 May 2026)
• anthropics/claude-for-legal on GitHub
• Claude for the legal industry: a practical deployment guide (PDF)
• Updates to Consumer Terms and Privacy Policy (Anthropic, 28 August 2025)
• Claude docs: Multilingual support
• HAQQ docs: Pseudonymisation
• HAQQ docs: Credits and usage
• Deusser et al., arXiv 2609.11335
• Claude vs HAQQ, feature by feature
• Claude Opus 5.5 for legal work: our benchmark
• Justinian®, HAQQ's legal AI engine