Phase 8 — Production Patterns: optimize, extend & ship

From "it works" to production. Phase 7 gave you the building blocks. This phase operationalizes them — make the loop cheap, give it new senses, and package it so a whole team can install it.

Executive summary

What this phase makes you able to do, and why it matters.

Once you can build the machinery (Phase 7), three production concerns decide whether it's actually usable at scale: cost, capability, and distribution. This phase covers the cheapest token optimization there is — prompt caching (reuse the stable prefix; reads cost ~0.1×) ¹; how to give an agent eyes and hands to verify work no unit test can — computer use & browser agents, with the safety caveats that come with reading untrusted screens ²; and how to bundle your skills, hooks, and subagents into one installable, versioned plugin a team shares via a marketplace ³. The throughline: a setup that works locally isn't done until it's cheap, capable, and shippable.

Prerequisite: Phase 7 — Advanced Patterns (the building blocks this phase operationalizes), plus Phase 2 (context as scarce) and Phase 3 (oracles).

Learning objectives

By the end of this phase you can: - Cut cost & latency with caching — structure prompts so the stable prefix is reused, and know what invalidates it. - Use a browser/UI agent as an oracle — drive visual self-verification for work with no unit test — safely. - Package & distribute a plugin — bundle skills + hooks + subagents into one versioned unit, shared via a marketplace.

---

The big idea (in one sentence)

A working agent setup isn't finished until it's cheap to run, able to see its own work, and easy for your team to install.

Lessons (one concept each)

#	Lesson	The one idea
1	Prompt & context caching	Cache the stable prefix (tools→system→context); reads cost ~0.1×. Keep volatile content last.
2	Computer use & browser agents	Eyes + hands: screenshot→act loop; a browser as a visual oracle — but untrusted screen = injection.
3	Plugins & marketplaces	Bundle skills+hooks+subagents+MCP into one installable, versioned unit; share via a marketplace.

---

Phase diagram

flowchart LR
    M["the machinery from Phase 7<br/>(skills · hooks · MCP · security · subagents)"]
    M --> O["OPTIMIZE — cache the<br/>stable prefix (L1)"]
    M --> E["EXTEND — give it eyes<br/>& hands (L2)"]
    M --> SH["SHIP — bundle & share<br/>as a plugin (L3)"]
    O --> P["agents that are cheap,<br/>capable, and shareable"]
    E --> P
    SH --> P

---

Cheatsheet

Key terms

Term	What people say	What it actually means
Prompt caching	"saving answers"	Reusing the processed prefix (tools→system→context); reads cost ~0.1×, output unchanged 1.
Computer use	"it can use my PC"	A tool to screenshot + click/type a screen in a reason→act loop; untrusted screen content can inject it 2.
Visual oracle	"screenshot testing"	Using a browser snapshot as the correctness check for UI that has no unit-test assertion (Phase 3) 2.
Plugin	"an extension"	One installable unit bundling skills+commands+hooks+subagents+MCP/LSP; shared via a marketplace 3.
Marketplace	"an app store"	A git-hosted catalog of plugins — add it, then /plugin install name@marketplace 3.

Agent translation (same idea, different homes)

Concept	Claude Code	Codex	Cursor
Prompt caching	automatic	automatic	automatic
Browser control	computer use · Playwright MCP	computer-use tool · Playwright MCP	Playwright MCP
Bundle & distribute	plugins + marketplaces	AGENTS.md in git	.cursor/rules in git

---

→ Check your understanding

---

← Phase 7 — Advanced Patterns · Curriculum home · more topics → Roadmap

---

1. Prompt caching — Anthropic ↩↩

2. Computer use tool — Anthropic ↩↩↩

3. Create plugins — Anthropic (Claude Code docs) ↩↩↩