A three-minute talking head from the new-model launch: less supervising, more directing. Treat Claude like a thought partner, give it goals it can verify, and be more ambitious.
Play video
Hey everyone, it’s Thariq from the Claude Code team, here to talk about our newest model in Claude Code.
PLAYMUTE00:00 / 03:00CCEXPAND
THE RIGHT WORK, NOT JUST THE WORK RIGHT
00:01Hey everyone, it’s Thariq from the Claude Code team, here to talk about our newest model in Claude Code.
00:08It is an incredible model, and it has really pushed us at Anthropic to rethink how we work with and prompt Claude.
00:14We used to need to check if Claude was doing the work right — breaking tasks into small chunks, double-checking its output, catching when it stopped early.
00:24Now I increasingly find myself checking if Claude is doing the right work. It can run for hours at a time, it tests its own work, and honestly it often produces better code than me. My job is more about direction and setup, not supervision.
00:43So here are the three ways I’ve changed how I work: treat Claude like a thought partner, give it goals and ways to verify them, and be more ambitious.
TREAT CLAUDE LIKE A THOUGHT PARTNER
00:00One of the failure modes I see is that I might not actually know what I want, or I might not know what is possible. By involving Claude early as I’m thinking it through, I can catch these problems before implementation.
00:15So I’d start with a small spec and ask Claude to interview me about the implementation before writing the final spec file. That builds my confidence that I know what I want.
PROMPT › on screen at 01:17
Interview me about the implementation before writing the final spec file.
00:25Or I might give it an idea, ask it to think of a few directions it could take it, and make some mockups in HTML for me to review.
PROMPT › on screen at 01:24
Think of a few directions you could take this, and make some mockups in HTML for me to review.
00:33When I’m ready to implement, I give it context, not just constraints, so it can help me actually accomplish my goal.
PROMPT › on screen at 01:39 — instead of “keep it simple, don’t over-engineer this”
This feature is an experiment. There’s a real chance we delete it in a month. Don’t build anything that would be painful to throw away.
00:49Giving it this context lets it catch things you might not even have thought of.
GIVE CLAUDE GOALS AND WAYS TO VERIFY THEM
00:00Once you know what you want — especially for an ambitious problem — give Claude goals and ways to verify them.
00:04Two features work well for this: /goal and workflows. Goals help Claude keep working until it’s done, and workflows help it parallelize and verify its work.
00:15So after I have a spec, I might tell Claude:
PROMPT › on screen at 02:11
Set a goal to implement the spec fully, then use a workflow to verify each part of the plan, and prepare a report on what was implemented and if anything differed.
00:28This lets Claude use its capabilities as creatively and thoughtfully as possible while making sure it’s building what you actually want.
BE MORE AMBITIOUS
00:00Finally: just be more ambitious. This model has pushed me to think outside the box with my work — for example, I edited this video with Claude.
00:11If there’s something you assumed LLMs couldn’t do, give it a chance. We really think this raises the bar on what’s possible. Try it out and let us know what you think.