Claude Code's /goal, Skills, and Cowork: the product surface that shipped while competitors talked
Felix Rieseberg, the engineer who built Claude Cowork, sits in front of his webcam demonstrating a $20 hardware button wired to physically approve Claude actions with a satisfying click. Matthew Wensing, VP of Product at Customer.io, runs 200 iterations on a single leadership deck to avoid what he calls "slop"—the confident-sounding output that executives filter out in the first paragraph. And across thousands of enterprise seats, PMs are still asking "Chat vs. Cowork vs. Code?" as their first question, not their tenth, because nobody told them `/goal` exists.
Three surfaces. Three unrecognizable workflows. One product. Zero shared map.
Anthropic shipped the complexity most competitors won't touch for another year. Chat for conversational queries, Cowork for file-system work with connectors and skills, Code for local development with multi-agent orchestration. Each surface unlocks something the others can't: Chat gives you quick answers, Cowork reorganizes your desktop invoices using hash functions to dedupe files, Code runs 21-agent dev teams that rewrite their own skills after every 15 tool calls. The feature set is extraordinary.
The data from Pawel Huryn's setup tells the real story. Seventy percent of his work happens in Dispatch plus Claude Code web sessions. Twenty-five percent in Code running locally. Five percent in Chat—and he's explicit about keeping Chat at 5 percent, treating it as a dead end for serious PM work. That's the power-user end state. They don't know the three surfaces exist as distinct tools. They don't know `/goal` is a feature. They open Chat because it's familiar, hit the limits Pawel describes—no continuity across devices, no file access, no persistence—and assume Claude just isn't for them.
The adoption gap isn't a bug in the rollout. It's structural. One user has been running skills 20-plus times every day since December, building self-improving loops where Claude rewrites workflows based on what worked in the last ten sessions. Another user, same product, same enterprise license, opens Chat once a week and asks it to summarize a PRD. The distance between those two users is wider than the distance between ChatGPT and no AI at all, and Anthropic hasn't built the bridge.
Felix's floor-plan demo shows what happens when someone understands the surface they're on. He drops a 2D floor plan into Cowork, asks it to build an interactive 3D walkthrough, and watches it generate a navigable model where he can move furniture around in real time. Then he connects his email as an inventory database—every purchase confirmation becomes a queryable record—and builds a promise-tracking system that scans his Twitter mentions for commitments and surfaces them in a live dashboard. Same session, wildly different outputs, all possible because he knows which surface does what.
Matthew's 200-iteration deck shows the opposite constraint. He's not avoiding Claude. He's managing it like a brilliant junior hire who races to the finish line before understanding the problem. Give it a goal and it sprints. It generates the one-shot alignment deck, compresses years of organizational nuance into seven slides, and hands you something that looks complete but fails the executive filter immediately. Matthew's solution: feed context in layers, kill the eager suggestions the moment they appear, and treat the session like a thinking partner instead of an output machine. Six hours, 400 employees waiting, and he walks in with a Q2 roadmap that matches the complexity leadership has been living with for months. But he had to know that workflow existed—and know it was different from the workflow Felix uses, and different again from what Pawel runs in Code.
The fragmentation isn't accidental. It's load-bearing. Each surface creates switching costs competitors can't replicate overnight. You can clone Chat in 90 days. You cannot clone a three-surface ecosystem where skills rewrite themselves, connectors pull live data from Gmail and Notion, and local agents orchestrate 21-role dev teams. OpenAI has ChatGPT. Google has Gemini. Both are single-surface products. Neither has shipping code that shows they understand the workflow divergence Anthropic is building toward.
But that divergence only works if users can navigate it. Right now they can't. The first question PMs ask isn't "How do I set up a self-improving knowledge system?" It's "Which tab do I open?" The documentation gap is real. Tutorials show you how to write a PRD in Chat. Almost nothing on how a VP of Product actually uses Code in depth, or when to choose Cowork over Code for the same task, or why keeping Chat at 5 percent is the directive that separates power users from everyone else.
The 70/25/5 split Pawel runs isn't discoverable through the UI. You learn it from a podcast, or a newsletter, or another PM who already climbed the curve. Skills are described as "the highest alpha technique in AI right now" by someone who's used them 20-plus times daily since December, but the average enterprise user doesn't know `/goal` exists. The Hermes framework reached 100K stars faster than LangChain or AutoGPT reached 100K—partly because it solves a security problem Windsurf had, but mostly because it writes its own skills and users can see the value immediately. Anthropic's skills do the same thing. The difference is Hermes ships with a 20-minute setup guide and a 30-day rollout plan. Anthropic's skills ship with a dropdown menu and no explanation of why you'd use them instead of custom instructions.
These aren't edge-case questions. They're the questions blocking adoption for the cohort that would extract the most value—PMs who aren't ex-engineers, leaders who don't have time to reverse-engineer the product from GitHub repos, teams that need a clear "do this, then this, then this" path and aren't getting one.
The strategic bet is that complexity is a moat. Every surface you add increases the switching cost. A user running 21 agents in Code, with skills that self-improve and connectors pulling live data, isn't moving to ChatGPT no matter how good GPT-5 is. The workflow is too embedded. The context is too deep. The agents have rewritten themselves four times and now finish tasks in eight minutes that used to take 20. You can't port that to a competitor in an afternoon.
The distance between 'Claude subscriber' and 'Claude power user' is the widest gap in the product, and it's not closing on its own.
But moats only work if users reach them. They try Chat, hit the limits, and churn before they learn Cowork exists. Or they open Cowork, build one workflow, don't understand skills, and plateau. Or they hear about Code, see the terminal commands, freeze because they've never touched a command line, and retreat to Notion AI instead. The distance between "Claude subscriber" and "Claude power user" is the widest gap in the product, and it's not closing on its own.
Anthropic has the surface area. They have the features. They have proof from real users—Felix's 3D walkthroughs, Matthew's 200-iteration decks, Pawel's 70/25/5 split, the PM running 21 agents who hasn't written production code in 15 years—that the workflows unlock 10x value when you know they exist. What they don't have is the forcing function that moves users from surface one to surface two to surface three without requiring them to read nine podcast transcripts and reverse-engineer a GitHub repo.
The window is short. Competitors are watching. They see the adoption numbers. They see the fragmentation. If Anthropic closes the onboarding gap—ships the in-product guide that shows you when to switch surfaces, builds the progressive disclosure that teaches `/goal` the first time you need it, creates the 20-minute setup flow that gets non-technical PMs to their first working agent—this becomes the product category everyone else chases for the next three years.
If they don't, the complexity that's currently a feature curdles into the thing that kills momentum. Users who can't navigate three surfaces pick the competitor with one. The power users stay. Everyone else churns. And the most defensible product surface in AI becomes the cautionary tale about shipping capabilities faster than you can teach people to use them.
The data is clear. The workflows are real. The question isn't whether the product works. It's whether Anthropic can teach it before confusion becomes the reason users leave.