Latest / AI Fire Daily / #504 Neil: Claude Code Slash Commands You Won't Find By Default Yet
Transcript
- 0:00You type a command into an AI, nothing happens.
- 0:03Beat. It's not broken. It's just waiting for
- 0:05you to teach it how to think. We really do treat
- 0:08it like static software. We assume we're just
- 0:10passive consumers. But you can actually rewrite.
- 0:13It's cognitive pathways. Welcome to the deep
- 0:15dive. Today we're exploring five custom slash
- 0:19commands. Yeah, these are built specifically
- 0:21for the Claude Code app. But the catch is you
- 0:24have to build them yourself. Right. We are your
- 0:27hosts, and we're looking at Plan, Clear, Goal,
- 0:30Rewind, and BW. These commands bridge a really
- 0:34vital gap. They bridge the gap between AI just
- 0:37doing things and AI doing things your way. But
- 0:40before we build them, we need to understand a
- 0:42common frustration. You watched a developer demo
- 0:44on YouTube? Oh, yeah. They attack a slash command.
- 0:47It runs perfectly. Right. You try it on your
- 0:49machine, and nothing happens. People assume their
- 0:52software is just outdated, or they missed a secret
- 0:55plugin. But let's look at what a slash command
- 0:57actually is. It starts with a forward slash.
- 1:00It triggers an action instantly. Exactly. But
- 1:03the secret here is it's not native software.
- 1:07It's literally just a plain markdown file. We
- 1:10call it a skill. And it lives locally on your
- 1:12machine. Yeah. Claude reads that exact file when
- 1:15you type the command. That is why lists vary
- 1:18between users. I mean, does this mean we are
- 1:20basically programming the AI's internal habits
- 1:23from scratch? Yes, absolutely. You are dictating
- 1:26its operational boundaries. You write them in
- 1:28plain English. So they were just text files directing
- 1:31the AI's behavior behind the scenes. Exactly,
- 1:33like stacking Lego blocks of data. So let's look
- 1:36at the first command, the plan command. Because
- 1:38Claude has a tendency to rush, right? Yeah, the
- 1:40problem is Claude starts coding instantly. On
- 1:43big tasks, it picks the wrong approach. It wastes
- 1:46so much time. It's like a driver accelerating
- 1:48before checking the map. That's a great way to
- 1:50put it. The planned skill fixes this. It forces
- 1:53Claude to stop entirely. It forces it to read
- 1:56the task first. Right. It breaks the task into
- 1:58steps, it lists the affected files, it flags
- 2:01risks, and then it waits for your approval. So
- 2:04like building an event signup page. Exactly.
- 2:06Using HTML and CSS, Claude stops and asks question.
- 2:10It asks if you need email validation. And it
- 2:13does this before writing a single line of code.
- 2:15Yeah. But exception here. You skip this for one
- 2:19line, one file fixes. Right. Why shouldn't we
- 2:21just type out these instructions manually every
- 2:24time? Well, because human discipline fades, you'll
- 2:26forget, and then the AI messes up. Typing it
- 2:29repeatedly wastes your time. This automation
- 2:31saves your sanity. Perfectly said. Once you finish
- 2:35that task, you have a new problem. The AI memory
- 2:38becomes a liability. We should quickly define
- 2:41context window here. The memory space where the
- 2:43AI reads your current... Chat history. Spot on.
- 2:47The problem is old context slows down responses.
- 2:50Claude tries to connect totally unrelated tasks.
- 2:52Like debugging a login form today, then switching
- 2:55to writing a course product description. Exactly.
- 2:57You do not want code in your marketing copy.
- 2:59The clear skill fixes this. It wipes the slate
- 3:02clean. It tells Claude to treat the prompt as
- 3:04brand new. Yeah. It ignores past files completely.
- 3:06But there's a crucial detail here. New commands
- 3:09only load when a session starts. So you have
- 3:11to start fresh. Right. You must open a new chat.
- 3:14Then the clear command will appear. Does the
- 3:17AI naturally realize when a topic has completely
- 3:21changed? Uh, no, not at all. It lacks that awareness.
- 3:25It just tries to blend everything. No, it assumes
- 3:27everything connects together unless you explicitly
- 3:29tell it otherwise. Sponsor. Beep. We're back.
- 3:35So we've started a clean task. But how do we
- 3:38stop the AI from getting lazy? Yeah, this is
- 3:41the huge issue. Instructions usually say what
- 3:43to do. They don't specify when it counts as done.
- 3:46So Claude just stops halfway through. Exactly.
- 3:48And it requires you to constantly reprompt it.
- 3:51The goal skill changes this. It establishes a
- 3:54clear finish line. It gives it a strict condition
- 3:56to meet. Right. Claude works through the steps.
- 3:58It checks the result against your condition.
- 4:00It does this after every single step. And it
- 4:02only stops when fully met. Yeah. Beat. Whoa.
- 4:06I mean, imagine it iterating through 50 steps
- 4:08overnight, perfectly checking its own work. Like
- 4:10building a landing page. Exactly. Where it done
- 4:13strictly means a hero section, one CTA, and mobile
- 4:16responsive. But wait, how does it actually know
- 4:19it hit the goal without our human eyes on it?
- 4:21It runs internal verification loops against your
- 4:24text. It reads its own output like a human reviewer.
- 4:29It constantly checks its own output against your
- 4:31strict definition of done. Exactly. But even
- 4:34with clear goals, things go wrong. Sometimes
- 4:37the AI takes a completely wrong turn. We need
- 4:40an undo button that works. Because Claude can
- 4:43make a really bad mid -task edit, and rolling
- 4:46back code by hand takes forever. I still wrestle
- 4:49with prompt drift myself. Sometimes I let it
- 4:52go too far before fixing it. We all do, man.
- 4:55That's why the rewind skill is amazing. It lists
- 4:57recent session checkpoints. With short little
- 4:59descriptive labels, right? Yeah. You pick one,
- 5:02and it rolls back both the conversation and the
- 5:04code. It takes everything back to that exact
- 5:06state. So if Claude restructures files in a way
- 5:09you didn't ask for? You just rewind to the checkpoint
- 5:11right before it happened. Does rewinding mean
- 5:13we also lose any genuinely good code it wrote
- 5:16after that checkpoint? Uh, yes, unfortunately.
- 5:19Rolling back wipes the slate clean moving forward.
- 5:21You have to act fast. Yes. Rolling back wipes
- 5:24everything after that point, so you must act
- 5:26quickly. Right. So say the task is flowing nicely
- 5:29now, but suddenly a random thought pops into
- 5:32your head. How do you ask? without derailing
- 5:35the AI. That's the problem. Asking a side question
- 5:38breaks the flow entirely. Claude mixes the answer
- 5:41into the main task's code. It pollutes the context
- 5:43window. Exactly. The B2 skill is the fix. It
- 5:47answers the question directly. It doesn't affect
- 5:49ongoing files at all. And then it returns focus
- 5:52to the main task. Yeah. Like, say you are coding
- 5:55a landing page. You ask for SEO image naming
- 5:58patterns. It isolates that query. Right. It answers
- 6:01it and acts like the interruption never even
- 6:03happened. What happens if you actually want the
- 6:05side question to influence the code? Oh, well,
- 6:07then just ask it as a normal prompt. Only use
- 6:09this for isolated thoughts. Then just ask it
- 6:12normally. Only use this command to keep tangents
- 6:14completely isolated. Exactly. So to recap our
- 6:18big idea today, these five tools, Plan, Clear,
- 6:23Goal, Rewind, and B2. They don't exist until
- 6:26you build them. Right. And you definitely don't
- 6:28need all five today. Just pick the one your workflow
- 6:30misses most. Save the markdown file and try it.
- 6:33Yeah, exactly. See how it changes your workflow.
- 6:36Thank you for listening today. We spend so much
- 6:38time marveling at what AI can do. We really do.
- 6:42But perhaps the real leap forward is realizing
- 6:44we have the power to shape exactly how it thinks.
- 6:47You are no longer just a user. You are the architect
- 6:50of your own AI. Outtio Music.