The Chatbot Era Is Over. Here Are the Three Skills That Replace Prompting.
Five independent signals converged this week on the same conclusion: the browser-tab AI model is ending. Ethan Mollick's Codex data, Karpathy on Claude Tag, and Anthropic's own numbers tell you what to practice next.
By Forge Team
If you are still running AI the way you did 12 months ago — type a question, copy the answer, paste it somewhere — you are using a workflow that the professionals getting the most from AI have already left behind. This week produced the clearest evidence yet of what they have moved to instead.
Five signals, one conclusion
On June 25, Ethan Mollick co-authored an OpenAI research paper (arXiv:2606.26959) using internal Codex data. Agentic AI users grew more than fivefold in the first half of 2026 — and the fastest growth was not among engineers. The median OpenAI legal employee now generates 13 times more monthly output tokens via Codex than in January. Researchers: 50 times more. Mollick's conclusion in his own words: "The chatbot era is over."
The same week, Andrej Karpathy called Anthropic's Claude Tag launch "the 3rd major redesign of LLM UIUX" — AI moving from a tab you open to a persistent Slack teammate with org-wide context, ready to act before you ask. Anthropic's Economic Index (June 26) showed 93% of Claude conversations now produce a concrete artifact. One-third of users surveyed expect AI to handle most of their work within 12 months. Google shipped computer use directly into Gemini 3.5 Flash (June 24), enabling agents that click screens and fill forms at roughly one-fifth the cost of GPT-5.5.
Five independent signals. Same week. Same direction.
What the shift actually demands from you
The chatbot taught you to prompt. What is replacing it requires three different skills.
Delegation. Scoping what an agent should do before it starts, not correcting what it did after. This means defining inputs, outputs, permitted actions, and the specific points where the agent must pause and confirm before continuing. If you cannot write a two-paragraph brief for an agent task, the agent is not ready to run on its own.
Supervision. Knowing when to intervene and when to let the agent work. Always-on agents like Claude Tag in Slack have ambient context your one-off prompts never did. An agent that reads your channels and drafts responses is useful. An agent that sends them without your review is a different proposition entirely — and the supervision design is what separates the two.
Workflow design. Building multi-step processes where AI handles execution and humans handle the judgment calls. Not "write this email" but "here is our response process for supplier complaints: step 1, flag and categorize; step 2, draft based on complaint type; step 3, human review before send." The professional who can design that process will get more from AI than the one who can only use it one prompt at a time.
Write the task brief for one repeating workflow — inputs, outputs, what the agent decides, and where it must pause and confirm before continuing.
Serena: from 40 manual summaries a day to one review queue
Serena manages client intake at a 90-person legal services firm. Her day used to start with 30 to 40 intake emails that needed to be read, summarized, categorized by matter type, and routed to the right attorney — all before 10 a.m. She had been using AI for the summarizing step only, one email at a time with a copy-paste prompt.
After reading Mollick's research this week, she redesigned the whole process. An agent now reads intake emails, assigns a matter category, drafts a three-sentence summary, and flags urgency level. The output drops into a Slack channel for her review. She confirms the routing and presses send. The agent never sends anything — her review step is non-negotiable.
Her throughput more than doubled. Her routing accuracy improved because she has time to actually check the agent's categorization instead of rushing through summaries. The delegation design took two hours. The supervision decision — what the agent can do without her, what requires her confirmation — took 30 minutes and was the most valuable thinking she did that week.
For one agent you use or plan to use, write down what it can do autonomously, what it must confirm before doing, and what should trigger an immediate stop.
What happens when you skip the design step
Marcus runs operations at a 15-person architecture practice. He tried deploying an agent to organize project files based on naming conventions. Three days in, it moved files that were live in client presentation folders, breaking shared links mid-project. His client noticed before he did.
The error was not the agent. It was the scope. Marcus had not written the one rule that would have caught it: "never move files referenced in active project folders." The scope document he now requires before any agent task includes one question that was missing: "What should this agent never do without asking me first?"
He now treats agent setup the same way he treats a new hire's first week: supervised on day one, spot-checked by day five, running independently on defined tasks by week two.
Map a repeating task, identify the steps that can run without you, and design the handoffs where human judgment takes back control.
The skill shift in one sentence
The chatbot era taught you to prompt. The agent era requires you to manage — which is harder, worth more, and something you can start practicing this week.
Like this post?
Get the next one in your inbox. Practical AI skills, no filler.