Lawyers don't distrust AI research because it isn't useful. They distrust it because a made-up citation looks exactly like a real one. Find that out in front of a judge, and no amount of time saved was worth it.

So the question isn't whether a companion can find the law. It's whether it works the way you already do. Jurisdiction settled first. Commentary, then primary law. Binding authority kept apart from persuasive. Every citation checked against the source, not a summary of it.

Strawberry works in your browser, so it can use the databases you already pay for, and follow a citation into the source instead of stopping at the search result.

What Strawberry can help you do

Not sure where to start? Tell your companion the question and the jurisdiction, and it will suggest the right one. See the skill

Answer a question about the law

A rule from the wrong jurisdiction is worse than no answer. It still looks like an answer. So nothing gets searched until that's settled:

  • You name the jurisdiction, and it gets on with it. If the facts point somewhere else, it says so.
  • You don't name one, and it asks first.
  • You're not sure which applies, and it lays out the plausible regimes and what decides between them: where the parties are, where the contract was formed, where the harm happened. Confirming which one governs is yours, or a qualified reviewer's, before anything rests on it.

From there it researches the way you were taught. Commentary first, to find the landscape and the authorities that actually control it. Then the statutes and cases that settle the question.

Everything comes back labeled: which jurisdiction, primary or secondary, binding on you or merely persuasive. A persuasive case from somewhere else can still be worth having on a novel question. It just arrives marked as one.

Your matter stays out of the search box, by the way. It searches the concept, not your client.

Understand what a new law requires

Something lands, and the questions are immediately practical. Does this apply to us? What do we have to do? By when? Most of that work is scope rather than summary. The thresholds and exemptions are what decide whether you're in or out.

The same instrument comes back pitched at whoever has to read it:

Written forWhat you get
A specialistFull detail, article-level references, no hand-holding
A clientPlain language, what they need to do, terms defined
A boardOne page, the headline risk, and the number

Binding law, guidance, and proposals stay clearly apart throughout. People act on proposals as though they were already in force, and that's worth heading off.

Compare the rules across your markets

One question across five regimes is really five questions. The usual shortcut is to research the first one properly and then go looking for its features in the others, which quietly reshapes one regime to look like another.

So each jurisdiction is researched on its own terms first, and kept apart while that happens. The point is that nothing from the German answer quietly colors the French one. Only once each stands on its own do the findings get compared.

Then the columns get normalised, because plenty of things look comparable and aren't:

  • Thresholds that are measured on different bases.
  • Penalties written as a fixed cap in one place and a percentage of turnover in another.
  • Dates that mean "in force" in one column and "applicable from" in the next.

Where a difference won't normalise, it stays visible instead of being tidied into a cell. What you usually want at the end is the strictest position on each row, because that's what a single global policy has to satisfy.

Know who you are dealing with

Plenty of legal work turns on the organization rather than the law. A counterparty before a deal, a supplier before you sign, the other side before you advise. That's company research with the sources kept attached, and it's the same companion doing it.

The seven ways AI legal research goes wrong

AI legal research fails in a small number of predictable ways. They're worth naming, because the workflow is built around heading off each one rather than hoping for the best.

How it goes wrongWhat stops it
Invented sources. The most dangerous one, because a fake citation looks exactly like a real one.Every source needs a URL or a database reference. Anything that can't be located is flagged as unverified, never quietly included.
Invented pinpoints. The case is real, the paragraph number isn't.Quotations get pulled from the actual text rather than written from memory. Page numbers that can't be confirmed are marked approximate.
Jurisdictional drift. A rule from one system presented as though it applies in another.Every rule is tagged with the jurisdiction it came from, in every output.
Outdated law. A statute since amended, a case since overruled.Statutes get checked against the current version and cases for whether they still stand. In fast-moving areas, anything more than a couple of years old is flagged for a manual look.
Persuasive authority dressed as binding.Each authority is labeled binding or persuasive for your jurisdiction, not left for you to infer.
A blog post cited as though it were the law.Secondary source claims get traced back to the statute, case, or regulation, and that primary source is what gets cited.
Confidentiality slipping into a search query.Client names, deal terms, and privileged facts never go into a search. It searches the concept instead.

You still verify. That doesn't change. The difference is that the draft arrives with the predictable traps already handled, so your check is a check rather than a rescue.

Build it around your practice area

Most of the value turns up on the tenth matter rather than the first. Once a way of working is settled, it can be saved: the jurisdictions you work in, the sources you trust, your citation system, the shape of output you actually want.

Teams that go further build a companion around the practice area itself. The workflows that cover most of your matters, plus the regulatory landscape, the enforcement history, and samples of past work product so it knows what good looks like in your firm. That is real investment up front, and it only makes sense where the same kind of matter comes round often enough to repay it.

For an area you track continuously, a Routine can watch agreed official sources and report only what has genuinely moved. Worth setting up once you know the sources, the cadence, and what counts as material.

Official Strawberry skill
Copy skill file Download skill file

Getting Started with Legal Research in Strawberry

Help the user complete one real piece of legal research, then show them the workflow behind it. This collection covers research, briefing, and comparison. It does not give legal advice, and it does not replace the user's professional judgment.

Two things are true of every workflow here, and they are the reason the rest is usable.

The companion produces the first draft and the source assembly. A qualified human verifies. Every citation must exist, support the proposition it is cited for, carry correct pinpoints, be current, and belong to the jurisdiction it is attributed to. Anything that cannot be verified is removed or explicitly flagged. Treat page references as unverified until checked; article and section references are structural and more reliable.

Confidential matter detail never enters a search query. No client or party names, deal terms, financial figures, or privileged facts. Use anonymized formulations. If the user pastes identifiable confidential material without having confirmed their confidentiality posture, flag it before processing, even mid-workflow.

Start with jurisdiction

Before any research, establish which jurisdiction's law applies. Ask when it is unspecified. When the user does not know, set out the plausible regimes and the factors that would decide between them, and treat that as a preliminary view for the user or a qualified legal reviewer to confirm rather than a settled answer. Never blend jurisdictions in one answer.

Choose the workflow

  • Answer a question: strawberry/legal/research-a-legal-question — secondary sources first, then primary law, with binding and persuasive authority kept separate.
  • Understand a new instrument: strawberry/legal/brief-a-regulation — scope, obligations, penalties, and dates, pitched at an expert, client, or board audience.
  • Compare several regimes: strawberry/legal/compare-across-jurisdictions — one like-for-like table plus what the differences mean in practice.

For background on an organization rather than the law — a counterparty, a supplier, an opposing party — use strawberry/research-analysis/research-a-company. To collect the same fields across many registers, filings, or case records, use strawberry/research-analysis/extract-web-data.

Where the browser matters here

Legal research lives behind subscriptions. Strawberry works in the browser, so it can use the databases and journals the user is already entitled to access, follow a citation from commentary into the primary source rather than stopping at a summary, and keep the citation and date attached to each finding. The user can watch it work and take over in a source when they want to read the passage themselves.

Keep what works

After a workflow proves useful, preserve the accepted jurisdictions, source hierarchy, citation system, evidence bar, and output format as a custom skill. For an area the user tracks continuously, a Routine can watch agreed official sources and report only material developments — once the sources, the definition of material, the cadence, and the review point are settled.