The bait, then the rug-pull.
Claire Vaux opens cold from her studio: she just got back from Code with Claude — Anthropic's first developer event — and has five things to show you in under ten minutes. No drama, no hype reel; she just sits down and starts naming what shipped, what's research preview, and what she'd actually build with each. The title screams power-up, but the video is closer to a calm field report from someone who reads changelogs for fun.
What the video promised.
stated at 00:14“I'm gonna walk you through five things that launched today, how they work, what they are, and what I might build with them. We're gonna keep it under ten minutes.”delivered at 10:00
Where the time goes.

01 · Cold open
Claire frames the video: she attended Code with Claude, will walk through five things in under ten minutes.

02 · Promise of five things
Sets the scope: what they are, how they work, what she'd build with each. Promises to keep it under ten minutes.

03 · Routines — live demo
Walks through the new Routines surface in the Claude Code app. Clicks New routine, names it Weekly newsletter, points it at changelog.md, sets Mondays at 6AM, picks a folder. Compares the feature to OpenCLAW crons.

04 · Routines — three trigger types
Slide breakdown: cron, HTTP webhook, GitHub webhook. Connectors (Slack, GitHub) come along. Runs local or in the cloud. Example use case: weekly PRD audit posted to team channel.

05 · Outcomes — rubric-graded agents
Anthropic's answer to OpenAI Codex 'slash goal'. Define what done looks like via a markdown rubric, the agent self-grades and iterates up to 20 times. Walks through a ChatPRD example: a rubric-graded 'Ship-ready PRD' mode.

06 · Multi-agent — coordinator + delegates
API lets you define a multi-agent team that shares one container and file system, up to 25 concurrent. Explicit orchestrator and sub-agent hierarchy, per-agent toolsets. Example: a PRD orchestrator with strategy, critic, and eng-review sub-agents working in parallel.

07 · Dreams — agent memory primitive
Memory framed as just writing markdown to disk. Most harnesses do it on a session-close hook. Dreams is different — it's an explicit API call to take N past sessions, review them, and surface what to commit to long-term memory. In research preview; she doesn't have access yet. Side-note about wanting a 'forgetting' counterpart.

08 · Usage limits doubled
Five-hour limits doubled across Pro, Max, Team, and seat-based Enterprise. Peak hours going away for Pro and Max. Opus API rate limits going up.

09 · Recap + SpaceX aside
Other stuff she didn't cover (data centers in space, SpaceX partnership). Re-summarizes the five practical takeaways. Lands on: none of it was mind-blowing, but it's all practical, and Anthropic is positioning as the agent platform of choice.

10 · Outro + show CTA
Bye + standard pod outro: like, subscribe, comment, find on Apple Podcasts / Spotify, see all episodes at howiaipod.com.
Visual structure at a glance.
Named ideas worth stealing.
Three trigger types for Routines
- cron (schedule)
- HTTP webhook
- GitHub webhook
Routines in Claude Code can be fired three ways and run either locally on your laptop or in the cloud, with connectors (Slack, GitHub) attached.
Outcomes loop (Anthropic's Ralph-loop equivalent)
- markdown rubric
- grader
- up to 20 iterations
- result: satisfied | needs_revision | max_reached
Anchor an agent on a rubric, let a grader judge each iteration against that rubric, loop up to 20 times until the agent self-grades as satisfied. Anthropic's flavor of OpenAI Codex's slash-goal feature.
Coordinator + delegates pattern
- one orchestrator agent
- up to 25 sub-agents
- shared filesystem/container
- per-agent toolset
Programmatic multi-agent team where one master agent drives work and sub-agents each have their own tools but share state. Demo: PRD orchestrator with strategy, critic, and eng-review sub-agents.
Dreams memory loop
- take N past sessions
- review them asynchronously
- surface important memories
- write to disk
An explicit primitive to consolidate agent memory on demand rather than on every session-close hook. Async, runs up to 100 sessions per dream, watchable in real time.
Lines you could clip.
“Define a rubric, give the agent the task, let it bang its head against that at least 20 times till it gets it right.”
“I think we think a lot about agent memory, but not a lot about agent forgetting.”
“This is what we do when we dream. We go through our day, we review it silently, and then we decide what to commit to memory.”
“None of this was mind blowing, but I think it's all really practical.”
“Anthropic is trying to be the agent platform of choice for builders.”
How they spent the runtime.
Things they pointed at.
How they asked for the click.
“If you enjoyed the show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app.”
Soft, post-roll, after she's already said 'Bye' to camera. Standard podcast outro pattern — no hard pitch, just menu of where else to find the show. Mentions howiaipod.com at the very end.
Word for word.
Steal the format.
A confident, calm 'I was there, here's what shipped' field report beats a hype reel every time — especially for changelog-style content.
- Set the contract in the first 30 seconds: number of items + time cap. 'Five things, under ten minutes' is a promise the viewer can hold you to — and is the reason they don't bounce.
- Use a split-screen frame: bold slide on the left, your face PiP on the right. Costs nothing in shoot time, doubles the perceived production value, and lets you read code on screen without looking like a screen-share.
- Per item, hit the same three beats: what it is, how it works, what you'd build with it. Repetition is the format — viewers feel the rhythm and stay.
- Pair every abstract primitive with one concrete demo. She tied Routines to a weekly-newsletter cron she actually built on camera. Outcomes got a 'ship-ready PRD' example. The viewer always has somewhere to land.
- End on a hot-take, not a recap. 'None of this was mind-blowing, but it's all practical' lands a thesis viewers can repeat — way stickier than 'so those were the five things'.
- Don't oversell what you don't have access to. She flagged Dreams as research-preview and admitted she hasn't touched it. That credibility carries the rest of the video.
What this means for your Claude Code workflow.
Three of these you can use today, one is research preview, and one (the usage-limit doubling) you already benefit from automatically.
- Routines: Open the Claude Code app, click New routine, give it a cron and a folder, and have your weekly chores run themselves — newsletter drafts, changelog audits, PR-rubric checks, whatever you currently do manually every Monday morning.
- Outcomes: When you have a quality bar you can write down (a 'good PRD' rubric, a 'good blog post' rubric, a 'good sales page' rubric), use the Outcomes API so the agent can self-grade and iterate up to 20 times instead of you doing the loop by hand.
- Multi-agent: If you've been trying to one-prompt your way through a job that really wants a critic, a strategist, and a reviewer, switch to a coordinator-plus-delegates pattern — up to 25 sub-agents share the same filesystem.
- Dreams: When access opens up, use it to consolidate the last N sessions worth of context into clean memory files rather than writing to disk on every session-close hook.
- Usage limits: Your five-hour Pro/Max/Team limit is now doubled, and peak hours are gone on Pro and Max — fewer pauses mid-flow.



































































