Posts

Two panes side by side. Left, labeled 'before · 3,569 words', a dense wall of skill instructions. Right, labeled 'after · 1,711 words', the same skill cut in half, with an eval report underneath reading 'candidate vs source: every assertion matched — PASS' and a red line showing the deliberately-broken control run failing.

Jul 20, 2026 · AI, Claude Code, agents, tools

Your agents' skills are too long (mine were too)

Anthropic cut Claude Code's system prompt by 80% because smarter models need less direction. I'd been sensing the same thing about my own skills, so I worked with my agents to cut them down — gating every cut behind an eval instead of vibes.

The Trip Shaper playground: vibe, pace, beach-vs-city, luxury, region, budget, and flights dials down the left; a live 'beach reset' trip card on the right; a 'Send to Claude' dock along the bottom.

Jul 13, 2026 · AI, Claude Code, agents, tools

Surface makes Claude's playgrounds real

Anthropic's playground skill is one-way: you turn dials, it hands you a prompt to carry to an agent that builds the tool for real. Wire it to surface and the agent is already the backend — the mockup is the live tool, and real results come back into the page.

Two windows side by side. Left, labeled 'the agent · still working', a terminal where a background session finishes an overnight run — tickets closed through the night, one ask filed — then runs 'reach andrew', picks the imessage channel as quiet hours end, and shows 'sent'. Right, labeled 'reach · how it finds me', an iMessage window from the agent: 'Overnight run's done — 9 tickets landed, tests green' and 'One thing needs you when you're back: the deal-card layout. I put both options on a page:' followed by a link.

Jul 9, 2026 · AI, Claude Code, agents, tools

Giving your agents reach

reach is a skill that lets your agents contact you — iMessage, email, push, whatever you actually use — without shipping a single integration. Your agent sets up each channel itself, learns your routing preferences, and comes to you when it needs you.

Two terminal panes side by side. Left, labeled 'act · work it does on its own', runs `act ready` and lists a prioritized queue of unblocked tickets, then `act next`. Right, labeled 'ask · what it needs me for', runs `ask list --open` and shows three items waiting on me — two blockers (authorize a spend, sign in to a service) and one normal request (a deal-card layout decision) — with the note '3 open · everything else kept moving'.

Jun 29, 2026 · AI, Claude Code, agents, tools

Two trackers that keep work moving while I'm away

act is a durable, agent-first task tracker. ask is an inbox the agent fills with the things only I can do. Together — with surface and reach — they let me kick off work and leave, getting pulled back only when the agent actually needs me.

A Claude Code terminal on the left using the surface skill to build a tic-tac-toe page; a browser on the right shows the live board, the player's center X highlighted as the agent drains it to respond.

Jun 18, 2026 · AI, Claude Code, agents, tools

Enhancing the effectiveness of HTML for AI agents

Anthropic's Thariq made the case for agents that hand you HTML instead of markdown — and today they productized it as Artifacts in Claude Code. Surface is a parallel take on the same idea: HTML that takes your input back, and reaches you off of chat.