At a glance
WHAT IT’S REALLY ABOUT
Scaling Claude Code with context-aware plugins, hooks, and agent workflows
- Out-of-the-box Claude Code is fine for simple coding, but large-scale software engineering needs customization so the agent can access the same information sources and workflows as engineers.
- Customization needs fall into three buckets—access, knowledge, and tooling—so Claude can see the “why” (decisions, docs, chats) and get fast feedback (CI, linters, LSPs).
- Because context windows are effectively bounded and KV caching makes early-prompt changes expensive, teams must be intentional about what goes into the prompt and where stable vs. volatile info lives.
- Different plugin primitives scale differently: MCP and skills can bloat the system prompt, while hooks provide a near “zero token overhead” path by running scripts externally and injecting only when needed.
- Claude Code’s emerging workflow features (worktrees, session coloring/renaming, agents talking to each other, /loop, auto permissions, Agents view, remote control) enable asynchronous, parallel development but require better context-switching habits.
IDEAS WORTH REMEMBERING
5 ideasIf Claude can’t access what you access, it can’t truly pair with you.
High-quality engineering depends on decision context (threads, emails, design docs, runbooks) that isn’t in the repo; connect those sources so the agent can infer intent and constraints rather than guessing.
Use an “Alt+Tab audit” to find missing integrations.
Work a full day staying in Claude Code; every time you switch apps to copy/paste info, write it down and later add a connector (often via MCP for external systems or scripts/skills for internal ones).
Prioritize tighter feedback loops over “smarter models.”
Hooks that run linters, typecheckers, tests, or LSP checks act like agent “red squigglies,” nudging quality continuously and preventing drift without hard-blocking progress.
Treat context as a constrained package manager running on an Arduino.
With context size largely plateaued, you must “not pay for what you don’t use”: keep stable, shared instructions early and inject task-specific, volatile details late so they can be swapped cheaply.
KV caching makes prompt position a cost and latency decision.
Changing early tokens invalidates downstream cache and can be ~10× more expensive; design prompts so frequently changing content is near the end and reusable scaffolding stays fixed at the front.
WORDS WORTH SAVING
5 quotesIf there's one thesis of this whole talk that I want you to take away here, it's that if Claude can't do everything you can do, it can't do your job with you, right? Your job as a software engineer at this point is to make little clones of yourself, so you can scale up your abilities and scale up your work, um, in, you know, across many agents, right?
— Daisy Holman
If you try doing a full day of work without leaving the Claude Code terminal or the desktop or whatever you use, right? Every time you have to reach for another tool, every time you have to Alt + Tab to something else and copy paste into Claude, that's something Claude is missing.
— Daisy Holman
ICL is like a fancy word for when you want people to think you're smart, but you're actually just talking about text files.
— Daisy Holman
The fastest way to make your agent better at your code base isn't a smarter model, it's a tighter feedback loop.
— Daisy Holman
I like to say it's like trying to run npm on an Arduino, right? You've got a, a tiny bit of memory, and you've gotta figure out the very most important things to put in there, and you wanna put the smallest version of it that you can in there in order to leave enough room to, to do real work, right?
— Daisy Holman
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome