Ann Srivastava

Law firms are putting AI in the wrong place. Sullivan & Cromwell, Latham, Allen & Overy - every major firm is racing AI into legal research, drafting, and memos. That's exactly where hallucinations become malpractice. A single fabricated case citation has already sanctioned real lawyers (Mata v. Avianca, 2023 - the ChatGPT lawyers). A hallucinated statute in client advice is worse. Meanwhile the one place AI is genuinely safe - intake, qualification, and scheduling still runs on PDF questionnaires and paralegal phone tag at almost every firm in the country. So last night I built what lawyers should actually be building. A demo website for a fictional U.S. immigration firm - Sterling & Reed, lead partner Ann Sterling (all names are fictional). An AI intake concierge named Evelyn qualifies every prospect through 17 consultative questions, books the consultation on Ann's Calendly, and emails a two-page matter brief straight to her Gmail before she joins the call. No briefs. No citations. No advice. No hallucination surface. Any immigration lawyer on this app can replicate it. Here are the 12 exact prompts I used - copy-paste into Claude Code: ━━━━━━━━━━━━━━━━━━━━ 1/ BRAND FOUNDATION "Design a boutique U.S. immigration firm website. Fictional founding partner, two offices (NY + Miami). Palette: deep navy + bronze + warm paper. Fonts: Instrument Serif for display, Inter for body. Luxury + editorial - no generic templates, no blue-chip blue." 2/ HERO "Full-screen dark cinematic hero. Centered serif headline: 'Your immigration lawyer, already [prepared].' The last word cycles every 3.5s, character-by-character morph - rotating through prepared / briefed / engaged / on your side. Background: 6 US city night-skyline photos crossfading every 5s with Ken Burns drift. Horizon glow + starfield overlay." 3/ AI INTAKE CONCIERGE "Build Evelyn, an AI intake concierge. 17-turn immigration intake: greeting → visa pathway → citizenship + status + expiration → professional background → timeline → visa-specific qualifier (EB-5 capital, O-1 evidence, E-2 treaty, etc.) → source of funds / sponsor → prior visa history + derivatives → red flag on prior denials → red flag on arrests / overstays → biggest concern → referral source → name → email → WhatsApp → present 3 slots → booking confirmation. Voice: warm, consultative, never rushed. Frame red flags as 'no wrong answers - Ann prefers to know upfront.'" 4/ THINKING STATE "Before each Evelyn reply, show a thinking state. Spinning bronze ring + context-aware label per turn ('Identifying visa pathway...' / 'Cross-referencing denials database...' / 'Preparing brief to Ann...'). Then typing dots. Then the reply. Feels deliberate, not robotic." 5/ AGENT AVATAR "Evelyn's avatar: real photo of a professional woman in a circle. Bronze conic-gradient ring rotating around her, sonar pulse rings expanding outward, green live-status dot bottom-right. Three states synced to chat activity: idle (gentle breathing), thinking (faster pulse + bronze glow halo), speaking (bronze waveform bars below photo)." 6/ BOOKING - CALENDLY INTEGRATION "After intake completes, embed the firm's Calendly inline in the chat for slot selection. On confirmation, show an animated card: 30-particle bronze burst + 4 cascading checkmarks 300ms apart - Brief delivered to Ann's Gmail → Calendar dispatched via Calendly → WhatsApp queued → Prep checklist attached." 7/ HOW IT WORKS - SCROLL-PINNED "4-step section pinned with GSAP ScrollTrigger: 01 Intake, 02 Routing, 03 Consultation, 04 Engagement. Each step: custom animated SVG (chat dots pulsing / checkmarks drawing / calendar slot pulsing / signature stroke drawing itself). As you scroll, active step scales up + glows, inactive steps dim + blur. Bronze progress bar fills the bottom of the active step." 8/ LIVE STAT BAND "One horizontal line: '1,247 Matters filed | 38 Countries of origin | 97% Approval rate.' White italic Instrument Serif numbers, bronze vertical rules between. On scroll-in: scramble-resolve animation over 1.5s. First stat then becomes a live ticker - every 10-24s increments by 1 with champagne flash + floating '+1 EB-5' / '+1 O-1' / '+1 Family' badge (weighted random matter type)." 9/ BEFORE vs AFTER "Editorial band showing '21 days → 6 minutes.' Left: huge italic serif '21 days' with diagonal strike-through that draws in on scroll + five struck-through bullets (12-page PDF, five emails, paralegal screening, conflicts memo, partner hand-off). Arrow. Right: italic bronze '6 minutes' + five clean bullets. Below: live session clock + three real-time counters (briefs filed, conflict checks cleared, calendar holds reserved) ticking up while the visitor reads." 10/ EDITORIAL TESTIMONIAL "Pull quote block. 200px italic bronze opening mark (❝) fades in at 18% opacity. Two-line quote with 300ms staggered reveal. Bronze underline draws under emphasized phrases. Below: bronze divider + initials circle + name + verified green pill ('● Verified client · 2025'). Bronze corner brackets top-left and bottom-right." 11/ REPRESENTATIVE WORK "3 recent matters as a vertical bronze timeline. On scroll, the line draws top-to-bottom and marker dots pop in with staggered sonar rings. Per matter: visa tag (EB-5 / O-1 / E-2), matter number ('No. 1,247'), italicized key figures, green outcome pill (I-526E Approved / Premium Processing Approved / First-Interview Approved)." 12/ BLOG + CONTACT + FLOATING BUTTONS "3 blog cards (Instrument Serif italic titles, bronze gradient placeholders): EB-5, O-1, Family-based. Simple contact form: name + email + country of citizenship + visa type + note. Dark footer with both offices. Two floating FABs: WhatsApp bottom-right (green sonar pulse, pre-filled message) + music toggle bottom-left." ━━━━━━━━━━━━━━━━━━━━ Built entirely in Claude Code. No Cursor, no React boilerplate, no design team. The intake bot runs as a deterministic server flow - no AI inference during the conversation itself, which is why it can't hallucinate. Briefs pipe to Gmail. Consultations book through Calendly. Deployed on Vercel in 15 minutes. Every tool a lawyer needs for this is either free or already in the firm. Swap the fictional firm for your name, your practice areas, your matters - customize and you're live by the weekend. The AI sits in intake, not in your brief. No hallucination, no malpractice, no sanction risk. Just a qualified lead, a warmed prospect, and a partner who walks into the consultation already prepared.

My grandfather was the first lawyer in our family. He bought his first computer in 1998. He was 64. The firm had told him he needed it for the new state e-filing system. He stared at the box on his desk for two weeks before opening it. A junior associate finally walked him through it. He never used it the way it was meant to be used. He printed everything. He still drafted by hand. The computer sat on the corner of the desk for the next eight years, mostly off. His clients went to younger lawyers. He spent the last five years of his career watching his book of business shrink and not understanding why. My father is a lawyer too. He learned email in 2003. He resisted LinkedIn until 2014. He still does not have a website. He is technically excellent - a better lawyer than I will ever be and his work has dried up. His clients are not going to better lawyers. They are going to lawyers who answer email faster, send proposals the same day, and look like they exist on the internet. I am the third generation. I watched both of them lose work to a shift they decided was a fad. So I started paying attention to the next one. This year - AI. The lawyers in my circle are doing what my grandfather did with the computer. They have opened Claude or ChatGPT once. They use it like Google. Many tell me they will "get to it later." Later never comes. I have seen this movie twice in my family. But here is the part nobody on legal X is writing about. The lawyers who moved fastest are also the ones in the headlines. Nebraska just suspended an attorney for AI hallucinations. California charged three. Federal courts are writing AI restrictions into protective orders. Moving fast without discipline is not adoption. It is malpractice with a faster timeline. So the choice is not adopt-or-die. There are three lawyers I am watching right now. The first refuses to learn. Her phone is going quiet. The second is using AI like a hire that needs no supervision. Her phone is going to the bar. The third built one workflow over a weekend. Verified every output. Wrote down her firm's standards. Tested before she shipped. Her phone is the only one still ringing in three years. The third is the one to be. Three questions to know which group you are in. Can you name which AI tool your firm uses and which tier? Has anyone verified the last citation you filed? If a client subpoenaed your AI chat logs tomorrow, would they find privileged content? If you cannot answer all three, you are in one of the two losing groups. My grandfather did not lose because he was a bad lawyer. He lost because the world changed and he decided it was someone else's problem. I am not making that mistake. If you have read this far, neither are you.

Every solo lawyer I have talked to this week has the same problem: "I am redoing the same thing over and over. Different client, same contract. Different case, same analysis. Different prospect, same intake." You are not doing 100 different tasks a year. You are doing one task 100 times with 5 percent variation. The job is not to make AI draft the work. The job is to make AI run the workflow end to end. How to find your one - if you have never used AI before. You do not need to install anything yet. You do not need a tool, a tier, or a prompt. You need 30 minutes and a pen. Step 1. Open your last 60 days of billable time. Print it. Stare at it. Group every entry by category. Contract review. Intake. Drafting. Research. Email. Whatever categories make sense for your practice. Count the entries in each category. Step 2. Find the one task you do 30+ times in 60 days. It will not be the one that pays the most. It will be the one you do without thinking. The contract you redline. The matter you intake. The demand letter you draft. The client status update you send. The closing checklist you run. That is your one. Step 3. Score it on two questions. Does the task require your judgment, or does it require your time? If it is judgment - your unique reading of the case, the strategy call, the client conversation - leave it alone. AI is not the answer. If it is time - the same structure, the same fields, the same checklist - it is your automation candidate. Step 4. Describe the task in three sentences before you touch a tool. What goes in? (The inputs - a transcript, a document, a form.) What comes out? (The output - a draft letter, a summary, a populated checklist.) What must always be true of the output? (The rules - voice, citations, jurisdiction, format.) If you can write those three sentences, you have already mapped the workflow. If you cannot, that task is not ready to automate. Pick a different one. The math. By hand: 30 minutes per instance, 100 instances a year. 50 hours of your billable time. Owned workflow: 20 hours to build once (a weekend). 5 minutes per instance after. 8 hours of run time per year. Net leverage after year one: 42 hours a year. Every year. Compounding. Stop renting features. Start owning workflows. If you want help mapping yours - DM "workflow" or grab the calendar link in my bio. The first 30 minutes is free.

@BitGrateful I know 😀 I took it as an opportunity!

@helloparalegal For the record my answers were satirical

Ten questions every lawyer should be able to answer about their firm's AI use in 2026. 1. Which AI tools does your firm actually use, and which tier? 2. Where does your client data go when you use them? 3. Have you read the terms of service for those tools? 4. Who verifies the citations in any AI-generated draft before you file it? 5. Have you written down your firm's standards for AI use? 6. What happens to AI-generated outputs from a matter after you close it? 7. Do you tell your clients when AI is used on their matter? 8. What was the last AI tool you stopped using, and why? 9. Are you on any AI tier that trains on your inputs? 10. Have you ever caught a hallucinated citation in your own work? If you can answer all ten without flinching - you are running a 2026 practice. If you cannot - that is the diligence call you make this month. This is exactly what we do. We walk solos and small-firm lawyers through this list. Write the standards file. Audit the stack. Set up the verification protocols. Test the citation discipline. 14 days. End to end. 5 June slots open. DM "audit" or grab the link in my bio.

Thank you for the honest answers. Let me walk through what each one actually means, because the gap between these answers and a defensible 2026 AI practice is real. 1. "The Open one" is ChatGPT. The tier matters. ChatGPT Free/Plus/Team are Consumer Terms - OpenAI trains on your inputs by default unless you actively opted out. Check Settings → Data Controls today. 2. "Secure" on the screen refers to encryption in transit, not whether the model trains on your input. The terms tell you. The UI does not. 3. Asking the AI not to make mistakes is not verification. The Nebraska attorney suspended last month asked very similar things. Real verification means opening every cited case in Westlaw or CourtListener and confirming it exists. 4. Citations that "look real" are exactly the failure mode. Hallucinated cites are generated to look real. Mata v. Avianca had six fabricated cites that all looked perfectly formatted. Looking right is the most dangerous signal, not the safest. 5. "Too confidential" as a guideline that bends "if it saves time" is not a confidentiality policy. It is a confidentiality risk. Write the standards down. Circulate. 6. Closing the browser tab does not delete the data. It sits on OpenAI's servers under whatever retention policy applies to your tier. For Consumer terms with training opted in - up to five years. You have just described the default state of nearly every solo practice in 2026. The post was about closing this gap on purpose.

1) We use the Open one. I’m not sure which version. Whatever one opens when we log in. 2) I assume it stays private because the screen says “secure” somewhere, or at least it looks professional. 3) No, but I asked the AI not to make mistakes twice. 4) The AI usually puts the citations in, and they are real. The names are exactly what real cases look like. 5) Not formally, but everyone knows not to paste anything “too confidential,” unless it saves a lot of time. 6) Once the browser tab is closed it is gone. 4) I don't know what that means but everything has going great.

I did not say I tested it. I have not. The post was about what shipped, not a recommendation that anyone deploy it. The argument was structural. Carta absorbed a law firm. Innanen released an open-source agentic firm pattern. Together they are a market signal about where legal services are heading - paid and locked in on one side, free and open on the other. Lavern's traction may or may not hold. That does not change the structural point. The pattern exists now. Forks will follow. If you have benchmarks or real test data on Lavern specifically, I would genuinely like to see them. The repo is four days old - if it stays at zero in six weeks, the open-source half of the signal weakens. Until then it is news, not endorsement.

@helloparalegal 0 stars, 0 forks, 0 watchers. That's telling. Have you tested it? Know anyone who has? Zero benchmarks worrisome unless we have verified user feedback.

Two things happened in legal tech this month that almost nobody on legal X is connecting. May 13. Carta acquired Avantia, an alternative legal services provider with 200+ asset manager clients supporting $15 trillion in assets under management. Rebranded as Carta Law. AI-native by design. The first major financial-services platform to formally absorb a law firm. May 20. Antti Innanen launched Lavern. An open-source "agentic law firm." 67 specialist agents. 155,000 lines of code. Mandatory human gates. 10-pass verification loop. Apache 2.0. Different companies. Same signal. The next wave of legal tech is not a tool you bolt onto your firm. It is a firm. For BigLaw the message is: a private-capital platform just integrated an entire law firm into its product. Clients can now route legal work through the same dashboard they use for cap tables. The unbundling has begun. For solos the message is different. Lavern's source code is Apache 2.0. The 67-agent architecture, the verification gates, the workflow orchestration - all of it is publicly available. A solo who studies the Lavern repo for one weekend now has the architectural reference for what an AI-native practice actually looks like. The solos who deploy real agent workflows in 2026 are not just running better practices. They are becoming the small AI-native firms the next Carta-style platform will acquire. The consolidation wave runs through small firms with built workflows. Not through small firms with the same setup they had in 2022. If you are building toward this DM "audit." 2 May slots are open. Cap is real.

Lavern repo: github.com/AnttiHero/lave… Two things worth opening first: The /agents folder - see how the 67 specialist agents are organized by practice area. Each agent is a markdown file with system instructions. Same pattern as Claude Skills. The verification loop in /workflows - this is the 10-pass quality gate. Read this code if you build any agent that touches client work. It is the cleanest open-source reference I have seen for hallucination protection. Apache 2.0 means you can fork it, modify it, ship it inside a closed product. Most permissive license you can build on.

@helloparalegal Repositório da Lavern no GitHub?

Lavern repo: github.com/AnttiHero/lave… Two things worth opening first: The /agents folder - see how the 67 specialist agents are organized by practice area. Each agent is a markdown file with system instructions. Same pattern as Claude Skills. The verification loop in /workflows - this is the 10-pass quality gate. Read this code if you build any agent that touches client work. It is the cleanest open-source reference I have seen for hallucination protection. Apache 2.0 means you can fork it, modify it, ship it inside a closed product. Most permissive license you can build on.

@helloparalegal Repo link?

@ryancarson Thankyou!

@helloparalegal Love this

I have done 47 DMs with lawyers in the last 30 days. Three questions come up every single time. "Am I already too late?" "How do I tell my paralegal what is happening without scaring her?" "What is the smallest thing I can build this weekend that proves to my partners this is real?" Three months ago none of these questions existed. Today they are the entire conversation. The lawyers DM-ing me are not asking whether AI matters. They are asking how to move without breaking anything they have built. That is a different stage of the curve. Here is what I am telling them. You are not too late. The lawyer next door is not ahead of you. She is also DM-ing me. The market is not 18 months ahead of you. It is closer to 18 days. The window to start is wide open. The window to claim "I was early" closes in spring. Tell your paralegal first. Before the partners. Bring her into the build before anyone else. Show her the work she will not have to do anymore - the intake forms, the conflict checks, the deadline tracking, the discovery indexing. The grunt work she did not go to paralegal school to do. Frame it as her getting her job back. Not losing it. Every paralegal I have brought into a build becomes the evangelist inside the firm. The ones who resist are the ones who were not brought in. Build the one thing that eats two hours every single week. Not the biggest workflow. Not the deepest moat. The smallest, most repetitive thing you do without thinking about it. Intake. Conflict check. Demand letter template. Status update email. Weekly time entry. Pick one. Replace it. Run it for two weeks. Show your partner the hours back. That is the demo. Two hours, repeatable, in a workflow they recognize. That is what flips a skeptical senior partner. Not a 47-slide deck. Two hours. The lawyers asking these questions are already past the "is AI real" phase. They are at the "how do I move" phase. That is the phase where careers compound. If you are still asking whether to start, you are reading the wrong post. If you are asking how - DM me. I will tell you which two hours to start with.

Different axis than the rest of the list. On confidentiality, Lexis+ with Protégé is the opposite of the consumer-tier problem - no training on your data, AES-256, and BYOK as of this month for the AmLaw crowd. Genuinely Acceptable-tier. On value it's a 'depends': enterprise pricing, no public number, almost always layered onto a Lexis sub you already pay for. Research enough and it earns its keep; don't, and it's an expensive way to do what Westlaw already does for you. Worth noting it runs on the same Anthropic/OpenAI models underneath. The security is the contract, not the brand.

@helloparalegal Curious where Lexis would rank

Worst subscriptions you're still paying for: [Lawyer's edition - 2026] Acceptable • Claude Pro ($20/mo) • Microsoft 365 • LawPay • Adobe Acrobat • Calendly Average • Clio Manage • Westlaw Solo • QuickBooks Online • LinkedIn Premium • Loom Regrettable ↓↓ • The $2,500/mo marketing agency posting generic LLC carousels • The answering service at $300/mo. Used twice this year. • Any legal AI tool that's a $200/mo wrapper on a $20 Claude account • The practice management software you migrated from in 2023. Still active. • That CLE platform you signed up for once. Last login: 2024. Which one's in YOUR Regrettable tier? Be honest. I won't tell your firm.

Right on the bottom line. Two precisions, because the precise version lands harder: 1. Not a new 'no privacy in AI chats' rule - just plain civil discovery. Relevant + no reasonable precautions = reachable, like any email. Privilege survives reasonable safeguards, which is the whole case for enterprise terms. 2. The 'preserve everything forever' order was OpenAI in the NYT MDL, and the going-forward part was lifted Oct 2025. What stuck: already-frozen logs, flagged accounts, and a 20M-log sample now headed to plaintiffs. The detail that proves your point: Enterprise + ZDR were excluded throughout. The contract is the protection.

No lawyer should be using $20/month Claude or ChatGPT subscriptions. Consumer tiers come with no Business Associate Agreement, no Zero Data Retention, and no contractual confidentiality protections. The provider's terms let them store, review, and in some cases train on whatever you paste in. The moment client facts go into that box, you've disclosed confidential information to a third party with rights to keep it. That puts your duty of confidentiality, and potentially attorney-client privilege itself, at serious risk. It gets worse. Your entire chat history sitting on the vendor's servers is almost certainly discoverable. Courts have already started ruling that there is no reasonable expectation of privacy in conversations handed over to a third-party AI provider, which means opposing counsel can subpoena it and the provider can be compelled to produce it. Some providers are currently under standing court orders to preserve every chat indefinitely, even ones users tried to delete. If you're going to use these tools in practice, you need an enterprise agreement with a signed DPA, written Zero Data Retention, and a no-training commitment. Otherwise you're outsourcing your clients' confidences, and your discovery exposure, to a vendor's marketing page.

The rule I give solos: Pro for admin, business tier for legal. Marketing copy, calendar logic, internal SOPs, summarizing a published opinion - Pro's fine. Client files, privileged facts, anything with a client's name attached - that goes to a tier that won't train on it and will sign a BAA. Match the plan to the data, not the task.

@helloparalegal Except that Claude Pro is NOT acceptable for lawyers. Friendly reminder: hoaglaw.ai/blog/kjedi-a-l…

Fair. Trust accounting is the part Clio earns its keep on. State-compliant IOLTA, three-way reconciliation, no commingling risk. If your practice runs heavy retainers, the $139 justifies itself on that line alone. If your trust activity is light, the $139 is mostly paying for the rest of the bundle. That is where it lands in Average for me.

@helloparalegal CLIO does credit card trust payments well

Save this if you are deploying anything in Obsidian this quarter. Most lawyers using Obsidian + AI are running one agent at a time. You type a question. The agent searches your vault. It writes back. You decide what to do next. That is the same shape as a Google search, just on your own files. Lexis just shipped something different inside their Protégé product. Three agents, not one. → A Planner that breaks the problem into steps. → A Research agent that pulls from the vault. → A Reflection agent that audits the output before the lawyer sees it. The lawyer never sees the back-and-forth between them. They see the finished memo. This is the architecture every litigation vault is moving toward in the next twelve months. Here is what it looks like for a solo: → Planner reads the matter file. Decides what the deposition outline needs to cover. → Research walks the vault. Pulls every prior deposition, every witness statement, every cross-examination outline from related matters. → Reflection checks the draft against the firm's playbook. Flags missing topics, weak inferences, hallucinations, unverified citations. Only then does the lawyer see the draft. One vault. Three roles. Same morning. Here is how a solo lawyer actually builds this. You do not need to write code. You do not need an IT team. You need a weekend. Step 1. Build the vault. A folder per matter. Inside each matter: Summary, Pleadings, Discovery, Depositions, Witnesses, Research, Strategy. Cross-matter folders for Witnesses, Experts, and Opposing Counsel - backlinked across cases. Step 2. Set up three Claude Projects. One Project per agent. Save these instructions exactly. Each Project becomes one role. PLANNER Project - system instructions: "You break legal questions into research subtasks. You do not answer the question. You only return the list of subtasks the Research agent needs to complete. Each subtask must specify what to look for, where in the vault to look, and what format to return." RESEARCH Project - system instructions: "You read the matter files in the attached vault folder. You pull facts, citations, witnesses, and prior statements relevant to the subtask. You return only what is in the files. If you cannot find something, say so. Do not invent." REFLECTION Project - system instructions: "You audit the draft against the firm's playbook file. Flag missing topics, weak inferences, hallucinations, unverified citations, and any factual claim not supported by the vault. You do not write the draft. You only audit." Step 3. Run them in sequence. Planner's output → Research. Research's output → Reflection. Reflection's output → yourself. Three handoffs. One workflow. Total time from question to first draft: about 30 minutes for a deposition outline that used to take half a day. Step 4. Wire it together when ready. Once the manual sequence proves itself, graduate to Claude Code with the vault as the working directory. The three agents become subagents. The handoffs happen automatically. You see only the finished draft. That is the build path. Manual today. Automated by the end of the quarter. The Obsidian vault is no longer just storage. It is the substrate the multi-agent stack runs on. If you are deploying agents one at a time, you are running 2024 architecture. The shift to multi-agent on your own files is the entire next twelve months of legal-AI work. If you know a solo lawyer building something in Obsidian - send this to that person. The three prompts above are the cleanest starting point I have seen for the architecture every firm is about to need.

Maya Reyes is a solo employment lawyer in Sacramento. She is 46, has two kids in elementary school, and left her 14-attorney firm two years ago because they would not make her partner. Her billables were "inconsistent" - which was code for the fact that she left every day at 5 to pick up her kids. She built her own practice in 18 months. She is profitable now. She has one paralegal, Jacqui, who is sharp and overworked. Between them they handle about nine new clients a month - wage-and-hour, discrimination, retaliation. Standard plaintiff-side mix. In February she signed up for Claude Pro. $20 a month. A former colleague had told her it would change everything. She tried it twice. Both times she typed something like "summarize this contract" or "draft a demand letter." The output was generic. She closed the tab and figured AI was a trend that was not for her. Six weeks later she was about to cancel the subscription. Then a friend showed her how she was actually using it. What followed is the playbook. Five features. Six weeks. The order matters. Week one. Prompts have five parts, not one. Context. Task. Standards. Tone. Format. The "summarize this contract" prompt became: "You are reviewing a commercial lease for a mid-sized retail tenant. Identify clauses that deviate from market-standard tenant-favorable positions. Flag anything that exceeds California retail-lease norms. Write for a senior partner presenting to a client. Format as a table with clause, position, market norm, risk level." Same contract. Same Claude. Completely different output. Maya stopped getting walls of text. She started getting tables that Jacqui could put straight into a client memo. Week two. Projects hold context across conversations. She had been re-uploading the same client documents every morning. Every new chat started from zero. She thought this was just how the tool worked. A Project is a workspace per client matter. Upload the documents once. Set custom instructions once - plaintiff, age 52, retaliation claim, California, flag anything affecting damages calculations. Every new conversation in that Project inherits the context. Jacqui started conversations in the same Project. Same documents. Same instructions. Maya stopped finding herself re-explaining the case to her own paralegal. Week three. Skills are plain text files. When she had a process she ran over and over - intake screening, demand letter drafting, wage-and-hour calculation - she stopped pasting the same checklist into every prompt. She wrote her intake screening playbook into a markdown file. Threshold issues to flag. Documents to request. Statute of limitations alerts. The standard questions she always asks plaintiffs in her first call. Now when she types "screen this intake," Claude applies her firm's standards automatically. The Skill is portable. Jacqui uses it. The file lives in Dropbox. Maya's intake process stopped living in her head. The first time she got sick for a week and Jacqui kept the funnel moving without her - that was the moment she realized what she had built. Week four. Co-work runs multi-step jobs. When the work required more than one step - read three offer letters, compare them, build a wage-and-hour calculation, flag retaliation patterns across the documents - she stopped doing it in back-and-forth chat. Co-work creates a plan. She approves the plan. Claude executes. She gets a spreadsheet at the end, not a wall of text. A comparison that used to eat Jacqui's afternoon ran in under four minutes. Week five. The Anthropic legal plugin ties it together. The plugin gave her slash commands built for legal workflows. /review contract. /triage NDA. /brief daily. The slash commands she invoked already knew her firm's positions, because they read her Skill file. The plugin and the Skill talked to each other. By the end of week five Maya stopped checking her email at 9 PM. The Skill ran the morning intake screen. Co-work built the comparison spreadsheets. The Project held every active matter's context. The prompts produced output Jacqui could use directly. Six weeks ago Maya Reyes was paying $20 a month for a wall-of-text generator and getting ready to cancel. Today she is profitable, home for dinner, and the firm runs whether she is logged in or not. The difference was not the model. It was knowing which five features to use, and in what order to learn them. Most lawyers I talk to are still stuck at week one. A note on Maya: composite. Names and identifying details changed. The arc is drawn from three solos I have worked with this year. The "first time she got sick and the firm kept running" moment in particular has shown up in every single one of them.

Griffin just told every BigLaw managing partner what they will be saying out loud in 90 days. 100 free legal agents shipped last week. Apache 2.0. Quinn Emanuel and Freshfields already on live matters. The trigger is pulled.