The crescendo begins — and a zero-shot demo
Welcome to the start of the crescendo. Today we tackle the big worry — large codebases — and then add some “spice”. But first, a treat: the remote Sprite from yesterday kept going, and produced something that ran perfectly on the very first try.
This is week three, day three — a core learning day. Originally it was going to be a whole day on large-team work, but we have already covered the foundations, so we can move quickly through the “vegetables” and then bring on some fun. Let’s start by digesting what we have eaten.
01A quick recap (digesting the feast)
The pro toolkit, one more time, so the terms are second nature:
| Tool | In one line |
|---|---|
| Slash commands | Configurable, but mostly replaced by skills these days. |
| Skills | The universal way to give agents functionality — consistent across Claude Code, Codex, OpenCode. Taking over from MCP. |
| Multi-agents | Many Claudes at once — locally, or in the cloud (GitHub @claude, or Sprites). |
| Sub-agents | Delegate one task to an isolated context (e.g. the built-in Explore agent) and get back just the result. |
| Hooks | Fire a command / prompt / sub-agent when something happens. Occasional; sometimes shipped in plugins. |
| Plugins & marketplaces | Package skills / sub-agents / commands and publish them for others to add. |
And yesterday’s sandboxing: the native /sandbox, managed Claude Code on the Web (ampersand, --remote, browser, mobile, and the delightful @claude issue), and third-party Sprites.dev — fast, stateful, and safe for full-YOLO because it is isolated.
02The Sprite that kept working
The remote box from 13.4 was too fun to switch off. At the end of the day the instructor asked it to do two more things — and it just… did them, unattended, in the cloud:
Consolidate the docs
Merge all the planning documentation into one clean summary document.
Build a demo
Create a terminal market-data demo to prove the simulator actually generates data — then push it.
Pull it locally
Back on the laptop: git pull to bring the new code down, ready to try.
03The live market-data demo
Running it locally — cd backend then uv run market-data-demo — produced a live, moving dashboard right in the terminal: tickers, prices flashing red and green, and text “sparklines” charting each stock’s movement.
04What “zero-shot” means
It worked on the first try — no “that’s broken, fix it” feedback loop. The demo was asked for once and simply ran. That is the bar coding agents are now clearing on well-scoped, from-scratch work.
Hold onto this feeling, because the next section is the counter-weight: agents shine on small, clean projects like this — the real skill is keeping that magic alive inside a massive team codebase. That is exactly what we tackle next.
✓ Key takeaways
- The pro toolkit in a sentence each: skills, multi-agents, sub-agents, hooks, plugins.
- Sandboxing recap: native
/sandbox, managed cloud (web/mobile/@claude), and Sprites.dev. - A remote Sprite consolidated docs and built a live demo — pulled locally, it ran zero-shot.
- Agents excel on small clean builds; the challenge (next) is doing this well on big codebases.