A team I can split on demand.

Haven't done a long journal entry in a while. Here's one.

First, something not super important but potentially useful: Claude 101.

This is Anthropic's own course. I'm bringing it up because so many people have zero idea why their results with LLMs are so bad. I've said it before: GIGO. Garbage in, garbage out. LLMs are more like a mirror of yourself, plus a highest-probability extraction machine; you can't feed a kid nothing but fast food and then blame the kid for getting fat.

anthropic.skilljar.com/claude-101 is the link to the Claude 101 course. Free to sign up, no weird fees.

The "Getting better results" page is literally teaching you how to get better output.

Next... health?

You can tell from the recent updates. Not wasting time rehashing it here.

Then... technical growth? LOL.

So, everything I do when creating bots, starting from the research phase (background, geography, culture, psychology, linguistics, etc.), has always been done in Claude Chat's Project feature. But Chat caps out at a 200K context window, and I have the awkward tendency to overshoot to somewhere between 210K and 260K. More importantly, Pro usage is nowhere near enough for me. I'm obsessive enough that I need Claude to repeatedly verify whether my ideas hold up, whether what we've researched is actually grounded in reality, and so on. On top of that there's the technical side: what holes exist in my bot definition if I write it this way, what natural LLM tendencies I need to reinforce against, that kind of thing.

So I ended up going Max 5x. (That's how a hundred dollars a month disappears, but my monthly token consumption is well past a thousand dollars' worth. Not a bad deal.)

After getting discharged from the hospital in March, I rested for nearly two weeks; my body was too oxygen-deprived to think, so I just played games. Mid-March, once I felt slightly less awful, I finally started looking into Claude Code and confirmed it was the best fit for my workflow. (Local file structure, lots of folders.) From that point on, Claude Code became my primary workspace for everything, bots included. I barely open the Chat-side projects anymore. When I do, it's for things unrelated to my pipeline, stuff that doesn't need constant read-write-edit cycles.

During this period I was still running Claude Code through Anthropic's own desktop app. I tried the CLI, took one look, gave up. But when Opus 4.7 launched in April and the desktop app dropped 4.6 support, I was forced to download Windows Terminal and run Claude Code through CLI. On the bright side, Windows Terminal is highly customizable. The reason I'd bailed on CLI before was the interface and font making me physically uncomfortable. (The customization settings were, of course, written by Claude after I described what I wanted verbally, then I copy-pasted the code. I cannot write a single line of code, and I have no intention of entering that field. So no, I don't hand-write code. Needed to make that clear.)

Then last week? Or the week before? GPT offered a free month of PLUS. For certain reasons I wanted to try GPT's Codex, which also runs through the terminal. Now every time I open Claude Code, a Codex instance comes with it.

Current workflow: Claude Code produces, Codex audits, and I'm the middle layer auditing both sides or copy-pasting between them. The updated master lorebook advanced script was also only completed because Codex helped identify the issue preventing it from running properly.

Today (yes, May 6th, this very day) I finished Anthropic's own courses:
Introduction to Subagents: subagent delegation, context management.
Introduction to Agent Skills: SKILL.md authoring, directory structure, team sharing.
Claude 101: the basics. I skipped all content and jumped straight to the quiz. 10/10. Done.

I finally have proper names for the things I've been doing all along, and actual shared vocabulary to describe them. Before this I'd been operating on instinct; I knew how to use the tools and how to get things done, but I didn't have the right terminology to actually show other people. (I'm using show deliberately, not teach. In Chinese the equivalent of "teach" carries a top-down connotation I don't like; show in English is more lateral, more direct. The Chinese word 展示 feels oddly stiff for this, so the concept doesn't map cleanly.)

My learning pattern: concepts came in and immediately collided with my existing workflow. The four application designs weren't post-course homework; they were simultaneous output generated while I was still mid-lesson. Concept acquisition and application mapping aren't two steps for me; they're one step. And I didn't do the binary of "study everything start to finish" versus "skip it entirely"; my approach to different courses on the same platform was completely different. Claude 101, one minute per page, grab the certificate. Subagents and Agent Skills, I stopped and actually digested. The filter wasn't course difficulty or the official recommended sequence; it was "does this concept fill a gap in my current workflow?" If it doesn't: certificate, move on. If it does: digest until I can output an application design. Also, I was running at least three Claude instances simultaneously, each with a different task boundary, manually routing information and deciding who sees what. My mental model is "a team I can split on demand." Though honestly this is just how I operate in general... from roleplay, to real-life work, to building bots and working with LLMs, I'm the one-woman crew: lighting, grip, editor, actor, director. Translated to workflow terms: I've hired myself a bunch of foremen, but the final creative decisions and the final sign-off are mine. The raw material the foremen work with has to come from me first before they can do anything.

Yeah. This is probably different from how most people work.

As for the bots, they should be more refined going forward, now that I'm using scripts. Of course, that's contingent on each character's script actually functioning correctly. And on JAI not suddenly throwing a tantrum and breaking something again.

It really is the classic scenario: most people think AI assistance means less work, more time saved. But for someone with standards, or let's be honest, a control freak, it's the opposite. AI hasn't made my workflow simpler or faster; it's turned what used to be a three-to-four-day bot into a one-to-two-week bot. Because now I can do more... things I used to be stuck on because I lacked the manual ability or the tools. Now that I have the tools, what excuse do I have not to do it right?

Still on the learning plan:
Anthropic's own courses: Introduction to Model Context Protocol (MCP fundamentals, building a server/client from scratch) and MCP: Advanced Topics (production-grade MCP, transport layer, sampling). Plus whatever other courses on their site I can sweep through for certificates. Grab them, dump them on LinkedIn, done. LOL.
LangChain's courses: Introduction to LangGraph (agent architecture, state graphs, tool calling, multi-agent systems; this one's the main course) and Introduction to LangChain (foundational framework concepts).