What 7,308 Agent Runs Taught Me About Writing Better Skills
SkillsBench tested 84 tasks across 7,308 agent runs and found curated skills boost performance by 16pp — but 19% of tasks got worse. It is a great backdrop to the things I have learnt too.
Two Files, One Workflow. No Code. Just Cowork.
I thought Cowork was a poor man's Claude Code. Building CoDraft changed that. Two text files in a folder — and when you open it in Cowork, you get a complete workflow. Here's how to build your own.
Budget 2026 Tells Lawyers to Use AI. But Are We in the Driver's Seat?
PM Wong named legal as one of the first professions needing AI transition support in Budget 2026. The government is putting real money behind this - 400% tax deductions, free premium tools, Champions of AI programme. But is this adoption infrastructure, or does it keep lawyers in the passenger seat?
My Voice Worked
This is a subscriber's only newsletter. There's a lot of ground to cover (Claude's
Legal AI's Real Value: Autonomy, Not Automation
Before AI, fear forced me to compromise. When a client's case didn't fit the template, I knew the optimal approach—but pursuing it meant hours of risk. AI didn't just save time. It gave me the autonomy to make the right decision, not just the safe one.
I Build Infrastructure. Jamie Vibe Codes Tools. Here's What I'm Missing.
Jamie Tso built 4 legal AI tools in 2-3 months using vibe coding. My redlines library took 3 years. The 30-90x velocity gap comes down to one question: am I building a tool or infrastructure? Most lawyers ask infrastructure questions for tool problems.
My 2026 Legal AI Predictions (From the Trenches, Not the Boardroom)
Stanford predicts AI's measurement era, Forbes predicts transformation, Above the Law warns about hallucinations. Nobody's predicting what will actually happen for resource-constrained practitioners. Here are five testable predictions with public accountability.
October 3 Changed Everything: How One $800 Fine Flipped Singapore's AI Coverage from Adoption to Accountability
On October 3, 2025, a Singapore lawyer paid $800 after an AI hallucination reached court. That single sanction flipped legal coverage from 89% adoption focus to 82% accountability warnings in eight weeks.
When Building Gets Cheap But Knowing Stays Expensive
In 2024, I spent hours crafting a 3-page prompt to generate an M&A term sheet for a legal tech competition. The result was a 4-page HTML document with timeline diagrams, color-coded risk tables (red/yellow/green), and professional typography that no standard Word template could match.
Why Prompt Engineering Felt Wrong (And What Skills Changed)
I sat in a standing-room-only prompt engineering workshop feeling anxious—not because I was behind, but because I don't use any of these frameworks. Turns out my instinct was right. That same month, the technology shifted from prompts to persistent agent skills
