AI for Lawyers

• Vibe coding: Describe what I want, Claude Code writes it. Review, iterate, deploy.
• Claude Code (Opus): my heaviest line item, on a paid Max plan (usage limits apply by plan tier). Handles 80% of my coding.
• ChatGPT Codex: Second opinion on architecture decisions.
• 2-4 hours per tool. Built over 2 years of consistent work.
• No coding background. I'm a lawyer who learned to prompt.

• Simple generators stopped selling. People use ChatGPT for basic NDA templates now.
• Complex generators still have demand. My Stripe demand letter with arbitration routing requires specific ToS knowledge ChatGPT doesn't have.
• Contract drafting from scratch is declining. Clients rarely ask me to draft ToS/PP from scratch anymore.

Most of my work is now reviewing people's ChatGPT first drafts.

Clients come with a GPT-generated contract asking: "Is this actually good? What did it miss? What's going to get me sued?"

That's exactly why I built Terms-Scanner - tools to validate AI-generated drafts.

• Regex patterns catch what GPT misses. 14 years of issues codified into patterns.
• 15-25 specific issues per scanner. Not "this looks risky" - specific issues like "indemnification is one-sided."
• Lawyers want legal intricacies. My explanations cite UCC, state law, specific principles.
• $0 per scan, instant results. No API calls. User pastes, JS runs, results appear.

• Step 1: Run scanner first (free, instant red flags)
• Step 2: Use Claude Sonnet for nuanced analysis (~$0.02/contract)
• Step 3: AI drafts redline language, I review and refine
• Result: 80% of initial review automated, I focus on what needs legal judgment

• Pick a contract type you review often (NDA, MSA, freelancer, lease)
• List the 6-8 issues you ALWAYS check for - that's your pattern list
• Write regex patterns to detect each issue
• Deploy as static HTML/JS - no backend needed
• Each scanner = one HTML file, takes 2-3 hours to build

• What works: Questionnaire-driven intake, JavaScript organizes into summary, AI drafts initial language
• Watch out: AI mixes up state requirements. CA community property, NY EPTL, FL homestead - all different
• Rule: Every document gets manual review for state-specific requirements

• What works: AI extracts info from docs, organizes for forms, drafts support letters
• Watch out: USCIS requirements are precise. Errors cause serious delays.
• Rule: AI drafts, human reviews everything. Strategy requires deep expertise.

• What works: Categorize docs, identify relevant emails, create privilege logs, research summaries
• Watch out: AI hallucinates case citations. Lawyers have been sanctioned for fake cases.
• Rule: Verify EVERY citation on Westlaw/Lexis. No exceptions.

• Criminal: Organize discovery, but jury selection, cross-exam, plea negotiations - all human intuition
• Family: Child custody involves reading people, not documents. Strategy is personal.
• PI: Intake forms and demand letters can be automated. Case value assessment is experience-based.
• Common thread: Where credibility and emotion matter, AI is a minor helper at best

This is where I spend. Opus-powered coding on a paid Max plan. I describe what I want, it writes. Review, iterate, deploy. The engine behind 5,000+ pages.

• Haiku 3.5: $0.80/1M in - quick categorization
• Sonnet 4: $3/1M in - main workhorse for analysis
• 10-page contract: ~$0.02 to analyze

Not free at this scale, but still incredibly cheap. Static HTML/JS files, no server to maintain. Near 100% uptime.

• Firebase: Real-time comments in workrooms
• Telegram Bot: Instant notifications when clients submit
• scheduler: Embedded scheduling on every page

Claude API with zero data retention enabled. Enterprise agreements. Never paste client-confidential data into a consumer AI account unless you have confirmed the account type, data controls, retention settings, and contract terms. Business and API products can be different; the risk is the specific account and configuration, not the brand name alone. Full breakdown of how this maps to CA RPC 1.1, 1.6, and 5.3 on my AI compliance attorney page.

No exceptions. The output is a starting point. I read, verify, refine. AI is a tool, not a replacement for judgment.

Mata v. Avianca: $5,000 sanction for fake AI cases. Lawyers have been publicly embarrassed and sanctioned. Check every citation. Every one.

I bill flat fees for most work. Client knows the cost upfront. AI makes me faster, they get better value. No ethical issue with efficiency. I also run this same stack-and-ethics review for other firms as a flat-fee AI implementation audit.

I tried AI for contract scanning first. Too slow, too expensive, inconsistent. Regex patterns are instant, free, 100% predictable. Save AI for novel analysis only.

All my tools are static HTML/JS files. Deploy to Cloudflare R2 and forget about infrastructure. Near 100% uptime, zero maintenance.

For demand letters, I don't use AI to generate text. I wrote templates, JavaScript fills variables. Consistent quality, instant delivery, $0 per letter.

Every scanner, calculator has scheduler embedded. User gets value, I get the lead. Convert while they're still engaged with your tool.