Intercom's top 5 AI users aren't engineers

Paul Adams, Intercom's Chief Product Officer, posted last week that how they build software is "unrecognisable vs 12 months ago." His engineer Brian Scanlan broke down the details: 90% of code changes at Intercom are now authored by Claude Code. They've built 13 plugins, over 100 skills, and automation workflows that turn it into their core engineering platform. And the five heaviest users of the system aren't engineers. They're design managers, customer support leads, and product managers.

This is how Intercom ships production software today. They gave Claude access to production data with real guardrails — the AI can read but not write to critical systems, every action is logged, and authentication runs through the same security their employees use. They built 30+ analytics workflows that sales reps, PMs, and data scientists rely on daily. And they built a feedback loop where every AI session gets analyzed automatically, problems get categorized, and fixes get filed so the system improves on its own.

Intercom's goal for Q2: every internal runbook followable by Claude.

We're hosting an official Claude Code workshop with Anthropic on April 10 in Midtown Manhattan. Half a day, 30 seats, hands-on. You bring a laptop, Anthropic engineers and AI strategists and builders from Tenex bring the expertise, and you leave with something you actually built — not a deck about what you could build someday.

Who it's for: Engineering leaders (EMs, VPs, CTOs, platform leads) and business leaders (CPOs, heads of digital transformation, senior AI sponsors). Up to 2 attendees per org.

What you'll do: Live demo of enterprise Claude Code workflows, then a guided build session where your agents write the code and you direct them. Show and tell at the end, networking lunch after.

Cost: Free. Application-based — we're keeping it small so everyone gets direct support.

Anthropic's Thariq Shaukat published the company's internal lessons on building skills last week: "A common misconception we hear is that skills are 'just markdown files.' The most interesting part of skills is that they're folders."

When your entire skill lives in a single file, Claude loads everything at once. Rules compete with examples. Context gets crowded. The longer the file, the more the output drifts. If you've ever had Claude nail one part of a task and completely lose the thread on another, this is probably why.

The fix from Anthropic's own practice is to use the folder structure so Claude loads information one piece at a time, pulling in only what it needs for each step. Say you have a skill that generates client reports. Instead of one massive file, you'd split it like this:

1. SKILL.md is the orchestrator. It contains no rules itself. It describes the workflow and points Claude to the right files at each step. "First, read the formatting rules. Then read the good example. Then generate the report." This is the only file Claude reads upfront.

2. instructions/ holds the actual rules. One file per concern. Formatting standards in one, tone guidelines in another, section-specific instructions in a third. They never compete because they're never loaded at the same time.

3. examples/ is where Claude learns taste. This is the part most people skip, and Anthropic says it's one of the most important things you can include. Good examples from real finished work show Claude what right looks like. Bad examples are just as valuable — they show Claude the specific patterns to avoid. Without both, Claude falls back on its defaults, which tend to drift over time.

4. eval/ is the quality gate. Anthropic recommends building a "gotchas" section from real mistakes Claude has made, then checking new output against it. You can set up simple pass/fail checklists, or go further and have Claude review its own work from different perspectives before calling it done.

One more tip most people miss: the description field at the top of your SKILL.md file isn't a summary for humans. It's how Claude decides whether to use the skill at all. Claude scans every skill's description at the start of a session to find relevant ones. Write it like a trigger condition ("use this when the user asks for a client report"), not a title.

Try it: If your skills are still single markdown files, paste this into Claude Code:

How to setup Claude Cowork so it’s actually useful — Full live walkthrough covering global instructions, connectors, folder structure, and three hands-on builds: a CEO coach, an AI advisory board, and a thought leadership pipeline. Includes the exact setup files. If you use Cowork and feel like you're barely scratching the surface, start here. Watch it (50 min)

Claude Can now use your computer — Full computer use just dropped in Cowork and Dispatch. Claude opens your apps, navigates your browser, fills in spreadsheets — anything you'd do at your desk. macOS only, research preview. Boris Cherny from Anthropic Labs (the team that shipped MCP, Skills, and Claude Code) says this is what they've been building toward since the early Sonnet 3.6 prototypes. Announcement

Anthropic's full skills playbook — Thariq's post that kicked off this week's conversation. 6.4M views. Covers nine types of skills (from API references to debugging runbooks), folder structure best practices, and how Anthropic manages their own internal skill library. If you're building with Claude Code, this is required reading. Read it

Cowork vs. Openclaw - when to use which — Solid 25-minute breakdown of how the two most popular AI agent tools compare, with setup tips and top use cases for each. If you've been wondering whether to go all-in on one or run both, this covers it. Watch it

Karpathy on No Priors — the full conversation — Beyond the "AI psychosis" framing, he walks through building a home automation agent named Dobby that controls his entire house via WhatsApp, auto-research that found optimizations he missed after two decades of ML, and why open source being 6-8 months behind the frontier is actually healthy. Watch it (1 hr)

Obsidian + Claude Code is the stack no one's talking about — 418K views and 5.6K bookmarks. Point Claude Code at your Obsidian vault and it answers from your own research instead of the internet. Your frameworks, your notes, your accumulated thinking — searchable and connected. Full guide

Lovable goes beyond apps — You can now build pitch decks, launch videos, PDF reports, and marketing assets inside Lovable, not just software. If you're in marketing or ops and haven't tried vibe coding yet, this might be your entry point. Announcement

"You're at the tip of the iceberg" — Alex Lieberman's thread on why drafting emails and smarter Googling barely scratches the surface of what AI can do now, and three concrete ways to start going deeper. Good starting point if you're still figuring out where AI fits into your actual workflow. Thread on X