Writing

Chat runs prompt-by-prompt — you steer every step. An agent has a goal, a process, and an exit condition; it runs on its own and stops when the goal is met. Subagents (Cowork and Claude Code): an orchestrator runs them in parallel, each works in its own context and reports back. Nice trick from Oleg Blinov — add a writer to the team and let the reviewers score 1–100 on their angles; the writer revises until every reviewer gives at least 90, with fresh reviewers each round so they don't see prior scores. Agent Teams (Claude Code, still experimental): agents are independent, see each other, message directly — continuous back-and-forth you watch live and can jump into. Subagents for tasks you'd split among separate people; Agent Teams for the ones you'd put in a room together to argue.

What does the EDPB say about automated decision-making after Schufa? Twenty minutes finding the right PDF, every time. Stack: a Google Drive folder mirrored as an Obsidian vault. Marker and Docxer plugins convert PDF and DOCX to clean markdown on right-click; Claude Desktop reads the whole base through the Drive MCP. The trick is the indexes — one per folder, one entry per document with a short summary. Claude scans the index, picks what's relevant, opens only those files. Same logic as a library catalog. Add a colleague to Drive and they have the same base, no onboarding. Public reference material only — for confidential documents, anonymize first or keep them outside the AI's reach.

10 contracts on your desk. Same counterparties across half of them. v1 anonymized one file at a time — Acme Corp became ORG_1 in one doc and could be ORG_3 in another, breaking cross-document analysis. v2 holds the matter, not the document: the same placeholder follows Acme Corp across all 10 files. Each anonymized file carries the session ID, so deanonymize one in 4 weeks and the real data comes back. Or export the matter as an encrypted archive — your colleague imports it with the passphrase and continues on their laptop. PII never moves between you and never reaches the API. Entity review now lives inside Claude Desktop instead of a separate browser. 95% recall on 50 English documents. One-drop install, no Python. Next: a CLI so PII Shield runs on a folder of contracts from any tool.

Honest question to legal colleagues: aren't you tired of writing boring memos? Tech founders open a 14-page .docx in Times New Roman, see numbered paragraphs and a nine-line "Executive Summary," and ask privately: could you just tell me what this means? With Claude's design tool I turned the Duke Memo into a web page in a single prompt — headings that breathe, sections you can navigate, citations that look like citations.

Two days after I shipped PII Shield (500+ reactions, 80+ comments) I spent a night building what people asked for. HITL review: a local web UI showing every detected entity color-coded in context — click to remove false positives, select text to add missed ones. Plus an audit log of every MCP call and an NER log of every detection decision, both on-disk under ~/.pii_shield/audit/.

An AI agent opened a case-law database, ran a search, screened 24 results, picked the four that mattered, followed their citation chains (ILVA A/S → Deutsche Wohnen → Nacionalinis → TR v Land Hessen), and delivered a structured digest. I made coffee. Not keyword matching — actual research behavior. Claude Max ($100/mo) holds the thread across 80-paragraph CJEU rulings without losing context.

I took our strongest memo and fed it to Claude's Skill creation. It parsed the .docx XML, figured out fonts, spacing, heading hierarchy, quote indentation, risk-assessment layout — turned it into a reusable SKILL.md. Now I throw raw notes from a call at it and get back a properly-formatted memo: Roboto 12pt, 11pt italic quotes, justified, single spacing, in my writing style. 20 minutes to set up. I still do the legal analysis. I just stopped doing the formatting.

I deployed an AI agent that finds trademark infringers and pirated app builds every week. App Store track: queries iTunes Search across countries, fuzzy-matches names and bundle IDs, narrows ~1,000 listings to ~50 candidates, runs icon vision on the high-risk ones. Piracy track: searches mod/cracked APK sites, parses download pages, collects evidence packages (screenshots, HTML, WHOIS). LLM scores risk and ships a prioritized Excel report. Single SKILL.md file — works on Claude or OpenClaw.