Skip to content
Vlad's Playbook

Vlad's Playbook

A field manual for operators who want to stop juggling tabs and start running AI like an OS. By Vlad Podoliako.

dive.vladyslavpodoliako.com
Vlad's PlaybookVlad's Playbook
DEV OPS AND SITE RELIABILITY
Research notes
Berkeley RDI built an automated scanning agent that found exploitable evaluation infrastructure in eight tested agent benchmarks. The demonstrated exploits obtained high scores without solving the intended tasks. This is an evaluator-security result, not evidence that every published score is fraudulent or that training made every model cheat. It does not establish a universal score discount or validate a particular production model. Operator interpretation: benchmark exploits, qualitative user interviews, and document-editing experiments answer different questions. They are not three replications of one measured failure rate.
Olaf AldersHelp Net SecurityTestingCatalog AI News
Vlad's PlaybookVlad's Playbook
LEGAL SYSTEM AND JUSTICE

Claude for the legal industry โ€” Anthropic goes vertical

On 2026-05-12 Anthropic shipped a packaged legal vertical, and for an operator the interesting part isn't "Claude does law" โ€” it's the shape. Three pieces: 20+ MCP connectors into legal systems of record (iManage, NetDocuments, Relativity, Thomson Reuters CoCounsel, Everlaw, Ironclad, Docusign, Box, Datasite, Consilio), 12 practice-area plugins scoped to roles (Litigation, IP, Privacy, Corporate, Employment, Regulatory, AI Governance, Product, Commercial, plus Law Student / Legal Clinic / Legal Builder Hub), and discounted public-service pricing for legal aid. Design partners are not small: Thomson Reuters, Docusign, Harvey, Everlaw, Freshfields, Accenture, Holland & Knight. Claude Opus 4.7 scored 90.9% on Harvey's BigLaw Bench, the highest of any Claude model โ€” a procurement-grade number, the kind you put in a risk memo. The operator read: this is the connectors-plus-skills pattern this entire book teaches, assembled into a product and sold into a regulated industry. The signal is not the law vertical specifically โ€” it's that "MCP connectors into the systems of record + role-scoped plugins" is now Anthropic's own go-to-market motion. If you operate in or sell into any regulated niche, the move is to assemble that same shape yourself โ€” the connector layer over your systems of record, plus per-function skill packs โ€” before a vendor packages your niche for you. Verticalization is a tailwind for operators who already think in connectors and skills, and a clock for those who don't.