Every Claude Code Concept Explained for Normal People

I have a confession: the first time I opened Claude Code, I closed it within ninety seconds.

A black window. A blinking cursor. No buttons. It looked like something out of a hacker movie, and I assumed it was "for developers only." I went back to the chat window I knew.

That was a mistake that cost me months. Because Claude Code is not a programming tool that happens to be useful for other things. It's an AI assistant that can actually do work on your computer — organize files, write reports, research topics, build spreadsheets, draft emails, maintain your content calendar — and it happens to be excellent at programming too.

The problem is that everything written about it is written by developers, for developers. The vocabulary alone is a wall: sessions, context windows, MCP, hooks, subagents, tokens. None of it is hard. All of it is badly explained.

So this is the article I wish someone had handed me: every Claude Code concept a normal person actually needs, explained in plain English, with a real business example and a copy-paste prompt for each one. Treat it as a free mini-course. Read it top to bottom, or jump to the concept that's confusing you today.

One promise before we start: zero hype. I'll tell you what each thing is, why you'd care, and — importantly — when you shouldn't bother.

---

Part 1: The Big Picture

1. What Claude Code actually is

Plain English: Claude Code is an AI assistant that can use your computer, not just talk to you about it.

Here's the difference that took me too long to understand. A chatbot is like a brilliant consultant on the phone: it can tell you how to reorganize your files, draft text for your report, describe steps to clean your spreadsheet. But you still do the clicking, copying, and pasting.

Claude Code is that same consultant, except now they're sitting at your desk with their hands on your keyboard. Ask it to "organize the 200 files in my Downloads folder into subfolders by client and year," and it doesn't explain how — it does it. It reads files, writes files, renames things, runs small programs, and reports back when it's done.

The analogy I use with clients: ChatGPT-style chat is a phone consultation. Claude Code is a temp worker who shows up at your office. Same brain, completely different usefulness.

Business example: A marketing manager I know receives monthly CSV exports from three ad platforms. She used to spend two hours merging them into one report. Now she drops the files in a folder and says "merge these into the usual monthly report." Four minutes.

Try this now:

I'm brand new to Claude Code and I'm not a programmer. Look at the files
in this folder, tell me what you see, and suggest three useful things
you could do for me here. Don't change anything yet.

2. The terminal (don't be scared of the black window)

Plain English: The terminal is just a way to talk to your computer with text instead of clicks.

That black window with the blinking cursor? It's not "hacking." It's the oldest, simplest interface there is: you type a sentence, the computer responds. Think of it as texting your computer instead of pointing at it.

You need to know embarrassingly little to use Claude Code in a terminal. Honestly, this is most of it:

You type	What happens
claude	Starts Claude Code in the current folder
Plain English sentences	Claude does what you asked
/help	Shows available commands
Esc key	Interrupts Claude mid-task
Ctrl+C (twice) or /exit	Quits

That's the whole "learn the terminal" curriculum for our purposes. You are not memorizing commands; you're typing English sentences to an assistant who happens to live in a text window.

Business example: An operations lead in a logistics company told me the terminal was her dealbreaker — until she realized she only ever types one command (claude) and then just… talks.

Try this now: open the terminal app (Terminal on Mac, or the Claude Code app — see the next concept), type claude, press Enter, and say:

Explain what folder we're currently in and what you can see from here,
like you're explaining it to someone who has never used a terminal.

3. The desktop and web apps (you may never need the terminal at all)

Plain English: Claude Code also comes in friendlier packaging — a desktop app and a version that runs in your browser.

If the terminal still feels hostile, good news: the same engine is available as a desktop application with normal buttons, panels, and a familiar chat layout, and there's a web version where Claude works on a computer in the cloud instead of yours. The concepts in this article apply to all of them — sessions, skills, permissions, all of it. The packaging changes; the machine inside doesn't.

My honest recommendation for non-programmers: start with the desktop app. It shows you what Claude is doing in a readable way, asks permission with clear buttons instead of cryptic prompts, and removes the "am I typing in the right place?" anxiety.

Business example: A solo consultant runs her whole proposal-drafting workflow in the desktop app. She's never opened a terminal in her life and doesn't plan to.

Try this now (in any version):

I want to build one repeatable weekly workflow with you. My role is
[your role]. My most repetitive weekly task is [task]. Interview me
with 5 questions, then propose how we'd automate it together.

4. The working folder (Claude's "office")

Plain English: Claude Code always works inside one folder on your computer — the folder you started it in. That folder is its office.

This is a small concept with big consequences. Claude sees the files in that folder, creates new files there, and treats it as the boundary of its world (it will ask before wandering outside). So the single best organizational habit you can build is: one folder per project or area of your business, and start Claude inside the right one.

Think of each folder as a client office. When the assistant is in the "Newsletter" office, everything on the desks is newsletter material. When it's in the "Q3 Planning" office, it's surrounded by Q3 documents. Context comes free with the room.

Business example: I keep folders like newsletter/, client-acme/, and research/. Each has its own instructions file (concept #8) and its own history. Claude in client-acme/ knows the client's tone, past deliverables, and current project without me re-explaining anything.

Try this now:

Create a sensible folder structure for my work as a [your role].
I mainly produce [reports / posts / proposals / analyses]. Make the
folders, add a short README in each explaining what belongs there.

---

Part 2: Conversations and Memory

5. Sessions

Plain English: A session is one continuous working conversation. Close it, and that particular conversation ends — but it isn't gone.

Every time you start Claude Code, you begin a session: an ongoing back-and-forth where Claude remembers everything said within that conversation. Yesterday's session? A different conversation. But — and this trips people up in a good way — you can resume past sessions. Ask to continue where you left off, and the whole previous conversation comes back, like reopening a meeting room where the whiteboard was never erased.

The mental model: sessions are meetings. Each meeting has its own discussion. You can walk back into an old meeting room, but by default each new claude start is a fresh meeting.

Business example: A researcher runs one long-lived session per literature review. Each morning she resumes it, says "where were we?", and Claude summarizes yesterday's progress and continues.

Try this now (at the end of any working session):

Before I go: summarize what we did this session, what's unfinished,
and what I should ask you first when I resume tomorrow. Save it to
a file called HANDOFF.md.

That file trick makes any session resumable even if you forget how to resume it properly.

6. The context window

Plain English: The context window is Claude's short-term memory — the amount of conversation it can "hold in mind" at once. It's big, but not infinite.

Imagine a whiteboard in your meeting room. Everything you discuss gets written on it. Claude can see the entire board at all times — that's why it remembers what you said an hour ago. But the board has edges. In a very long session, or one where Claude has read many large files, the board fills up. When it does, older material has to be condensed or erased to make room.

Symptoms of a full whiteboard: Claude starts forgetting earlier instructions, repeats questions, or gets vague. This isn't the AI "getting dumb" — it's the board running out of space.

The fix is hygiene, not heroics: keep sessions focused on one task, start fresh sessions for new topics, and put anything permanent into files (files don't expire; whiteboards do).

Business example: A content lead was frustrated that Claude "forgot" her style rules after long editing sessions. The rules lived only in chat — on the whiteboard. We moved them into a CLAUDE.md file (next concept), and the problem disappeared forever.

Try this now:

How full is your context right now? If we're past halfway, summarize
everything important into a file called NOTES.md so we can safely
start a fresh session without losing anything.

7. Compacting and clearing (/compact and /clear)

Plain English: /compact shrinks the conversation-so-far into a summary to free up space. /clear wipes the whiteboard completely for a fresh start.

When the whiteboard fills up, you have two janitorial options. /compact is the careful one: Claude summarizes everything so far and continues with the summary — like photographing the whiteboard, erasing it, and taping the photo in the corner. Details can get lost, but the plot survives. /clear is the ruthless one: total wipe, brand-new conversation, same folder.

My rule of thumb: /clear between unrelated tasks (finished the report, now doing email triage — clear). /compact mid-task when the conversation is long but you're not done. And anything truly important goes into a file before either.

Business example: An ops manager runs morning admin (invoices, inbox, calendar) as three tasks in one sitting, typing /clear between them. Each task gets a clean, sharp Claude instead of one dragging two unrelated conversations behind it.

Try this now (mid-long-session):

We're going to compact soon. First, write everything essential —
decisions made, current status, open items — into STATUS.md so
nothing important lives only in this conversation.

8. CLAUDE.md — the standing instructions file

Plain English: CLAUDE.md is a text file of permanent instructions that Claude automatically reads at the start of every session in that folder.

This is, no exaggeration, the highest-value concept in this entire article for a business user. Remember that sessions forget each other? CLAUDE.md is how you stop repeating yourself. It's the employee handbook sitting on the desk of Claude's office: every new session, Claude reads it before doing anything.

What goes in it? Anything you'd otherwise say every single time:

• "All client emails are formal; internal notes are casual."
• "Our prices are X; never quote below Y without flagging it."
• "Reports always follow this structure: summary, data, recommendation."
• "Never send or publish anything without showing me first."

It's a plain text file. You can write it yourself, or — better — have Claude write it from an interview.

Business example: My newsletter folder's CLAUDE.md contains my tone rules, banned words ("delve," "game-changer"), target length, and audience description. Every drafting session starts pre-briefed. Repeatable workflow, zero re-explaining. This is the exact mechanism behind my workflow template.

Try this now:

Interview me with 8 questions about my work, standards, tone, and
pet peeves. Then create a CLAUDE.md file in this folder so every
future session starts already knowing how I like things done.

9. Memory (what persists beyond the handbook)

Plain English: Beyond CLAUDE.md, Claude Code can keep its own notes between sessions — things it learned about you and your preferences while working.

Think of the difference between the employee handbook (CLAUDE.md — rules you wrote) and the assistant's personal notebook (memory — observations it accumulated). When you say "actually, always put dates in European format," a good assistant jots that down so tomorrow's session benefits without you editing any file. You can also explicitly tell it to remember things, ask what it remembers, and tell it to forget.

The practical takeaway: correct Claude out loud. "Remember that we call customers 'members,' not 'users'." Those corrections compound. After a few weeks, the assistant that shows up each morning is noticeably your assistant.

Business example: A recruiter corrected Claude twice about how candidate summaries should be formatted. Third session onward, every summary came out right without being asked.

Try this now:

What do you currently remember about me and my preferences from our
past sessions? List everything, and ask me what to keep, correct,
or forget.

---

Part 3: Commands, Skills, and Superpowers

10. Slash commands

Plain English: Slash commands are shortcuts — type / plus a word to trigger a built-in function instead of writing a sentence.

You've already met /help, /clear, and /compact. Slash commands are the speed-dial buttons on the office phone: /model switches which AI model you're using, /cost shows what you've spent, and so on. You could accomplish most of these with sentences; the slash version is just faster and unambiguous.

You don't need to memorize them. Type / and a menu appears. I cover the genuinely useful hidden ones in secret Claude Code commands.

Business example: A founder starts every working block with /cost (am I burning money?) and ends with /compact if she plans to continue later. Two commands, functioning routine.

Try this now: type /help, then ask in plain English:

Of all your slash commands, which five are most useful for a
non-programmer doing business work, and when exactly should I
use each one?

11. Custom slash commands

Plain English: You can create your own slash commands — one word that triggers a whole saved prompt.

Here's where "features" become "workflows." Any prompt you find yourself typing repeatedly can be saved as a personal command. Write your perfect weekly-report prompt once, save it as /weekly-report, and forever after, those thirteen characters replace three paragraphs of typing. It's the difference between explaining your coffee order every morning and being a regular — "the usual" — at your favorite café.

Custom commands are just small text files in a folder, and here's the beautiful part: Claude can create them for you. You never write them by hand.

Business example: A marketing team has /brief (turns a messy campaign idea into a structured brief), /social (turns any blog post into platform-specific social copy), and /compete (summarizes a competitor's positioning from their materials). New hires are productive on day one because the workflows are commands, not tribal knowledge.

Try this now:

Take the last big request I gave you, generalize it into a reusable
prompt, and save it as a custom slash command called /repeat-this.
Explain how I use it next time.

12. Skills (SKILL.md files)

Plain English: A skill is a folder containing instructions that teach Claude how to do one specific job your way — a recipe card it pulls out exactly when that job comes up.

If CLAUDE.md is the employee handbook (always-on, general), a skill is a recipe card in a box on the shelf (specific, pulled out only when needed). Each skill is a folder with a file named SKILL.md inside: the recipe. It says what the skill is for and gives step-by-step instructions. When your request matches — you ask for a cover letter, and there's a cover-letter skill — Claude reads that card and follows it.

Why not stuff everything into CLAUDE.md? The whiteboard, remember? CLAUDE.md is read every session and eats space every time. Skills load only when relevant. Handbook for the rules that always apply; recipe box for the procedures that sometimes apply. That separation is context hygiene — the same principle behind cutting your token usage in half and the whole discipline of context engineering.

Business example: My proposal skill encodes my whole method: qualifying questions to ask me, structure, pricing presentation rules, tone. "Draft a proposal for the Meridian project" now produces a 90%-done document in my voice. That's the differentiator I keep banging on about: features are nice, but encoding your method into a repeatable skill is where the leverage is.

Try this now:

Watch how I like [task] done based on our conversation so far. Create
a skill for it: a folder with a SKILL.md describing when to use it
and the exact step-by-step process, including questions to ask me
before starting.

13. Installing and downloading skills

Plain English: Because skills are just folders of text files, you can download ones other people made and drop them into your setup.

This is the part that surprises people most: there's a whole ecosystem of pre-made skills, and "installing" one is about as technical as saving an email attachment. A skill made by an expert encodes their method — download it, and Claude works their way for you. It's hiring their playbook without hiring them. Claude Code also has a built-in marketplace/plugin system, and the laziest correct approach is to simply ask Claude to install a skill for you — it can do it itself.

A few from my own library, all downloadable on this site: Superpowers (a famous collection of battle-tested workflows), Karpathy's guidelines (taste rules for AI-assisted work distilled from one of the field's best-known teachers), GSD Core (a get-stuff-done project execution system), and Skilld (a skill that helps you build better skills — yes, really). I keep a running countdown of the field in most popular AI coding skills — don't let "coding" scare you; most encode general working discipline.

Business example: An agency owner installed a research skill and a report-writing skill on Monday. By Friday her junior staff were producing senior-quality client research — the method was in the skills, not in anyone's head.

Try this now:

List the skills currently installed here. Then suggest three skills
that would help someone in [your role], and help me install or
create the most useful one.

14. MCP connectors

Plain English: MCP connectors plug Claude into your other tools — email, calendar, drive, databases, analytics — so it can work inside them, not just talk about them.

MCP stands for Model Context Protocol, which is a terrible name for a lovely idea: a universal power socket. Out of the box, Claude's hands reach your files. Each MCP connector extends those hands into one more system — Gmail, Google Drive, Notion, Slack, your SEO tools, whatever. Once connected, "check my inbox for unanswered client emails and draft replies" becomes a real instruction instead of a fantasy.

The analogy: your new assistant showed up with skills and a great attitude, but no keycards. Each MCP connector is a keycard to one more room in the building. And like keycards, you grant them deliberately — each connector asks for access, and you decide (see permissions, next).

Business example: I have an SEO data connector wired into Claude Code. "What questions are people asking around this topic, and which have weak answers?" is now a two-minute query instead of an afternoon of tab-juggling. Same pattern works for calendars ("find a slot for a 3-person meeting next week") and drives ("pull last quarter's proposals and list the pricing we quoted").

Try this now:

What MCP connectors do I currently have available? For someone in
[your role] using [your main tools], which 2–3 connectors would save
the most time, and what would a weekly workflow using them look like?

15. Permissions and safety

Plain English: Claude Code asks before doing anything consequential — and you control how much it's allowed to do without asking.

The single question every sensible person asks first: "So this AI can use my computer… what stops it from wrecking something?" Answer: a permission system. By default, Claude asks before editing files, running commands, or touching connected tools. You'll see clear requests: it wants to do X — allow once, always allow this kind of thing, or deny?

Think of a new employee's first week. Day one, they check before sending anything. Over time, you extend trust per task type: "you never need to ask before creating drafts" but "always ask before sending email." Claude works the same way, and the trust levels are yours to set. My honest advice for beginners: keep it chatty. Approving actions for a few days is how you learn what it actually does — that's a feature, not friction. Widen permissions only for actions you've watched it do correctly many times. And never grant blanket "do whatever you want" access on a computer holding things you can't afford to lose.

Business example: A bookkeeper allows Claude to freely read and summarize financial exports, but every file modification requires a click. The workflow is fast and auditable.

Try this now:

Explain your current permission settings in plain English. What can
you do without asking me, and what needs my approval? Recommend a
sensible setup for a cautious beginner.

16. Plan mode

Plain English: Plan mode makes Claude describe its full plan for your approval before touching anything.

Some tasks you want done; some you want proposed. Plan mode is the difference between telling a contractor "renovate the kitchen" and walking out for the day, versus "show me the plan, the materials, and the order of work first." In plan mode, Claude is read-only: it investigates, thinks, and presents a step-by-step plan. Nothing changes until you approve — and you can edit the plan before greenlighting.

Use it whenever the task is big, touches many files, or is hard to undo. Reorganizing your entire client folder structure? Plan mode. Fixing a typo? Just do it. This plan-first discipline is the beating heart of spec-driven development — and I promise the idea transfers beautifully from software to marketing plans, ops overhauls, and content strategy.

Business example: Before letting Claude restructure three years of chaotic project files, an ops manager ran it in plan mode. The plan revealed it would have merged two folders she needed kept separate. One sentence of feedback, revised plan, approved, executed. Zero surprises.

Try this now:

Enter plan mode. I want to reorganize this folder properly. Study
what's here and present a complete reorganization plan — what moves
where and why. Do not change anything until I approve.

17. Subagents

Plain English: Subagents are extra copies of Claude that your main Claude can send off to work on pieces of a task in parallel.

You're not talking to a lone worker; you're talking to a team lead who can staff up on demand. Ask for something big — "research these five competitors" — and Claude can dispatch five subagents, one per competitor, each working simultaneously in its own separate conversation, then merge their findings into one summary.

The quiet genius is the separate conversation part. Each subagent gets its own whiteboard, does its messy reading in its own room, and reports back only the conclusions. Your main conversation stays clean. It's delegation exactly as it works in a good office: the lead keeps the overview; the staff absorb the grunt work.

Business example: A market analysis that would have blown out one session's context — ten industry reports to digest — went to ten subagents. The main session received ten one-page summaries and synthesized them. Faster, and the "team lead" never got overwhelmed.

Try this now:

I need research on [topic] covering [aspect A], [aspect B], and
[aspect C]. Use separate subagents for each aspect in parallel,
then combine their findings into one summary with a comparison table.

18. Hooks

Plain English: Hooks are automatic rules that fire at specific moments — "whenever X happens, do Y" — enforced by the system itself, not by Claude remembering.

An instruction in CLAUDE.md is a request; Claude follows it the way a diligent human follows the handbook — reliably, but not mechanically. A hook is a rule wired into the building: whenever Claude finishes a task, play a notification sound; whenever it edits a file in the contracts folder, log it; block any action that touches the folder marked "originals." The badge reader doesn't decide whether to log your entry. It just logs it. Every time.

For most non-programmers, hooks are a "know it exists" concept rather than a week-one tool. When you catch yourself saying "I wish it always did X without fail," that's the hook-shaped hole. Setting one up involves a settings file — and yes, Claude will happily set up its own hooks if you describe what you want.

Business example: A consultancy uses a hook that flags any document Claude creates containing a price, routing it to a human review folder. Not because Claude misbehaves — because "pricing gets human eyes, no exceptions" should be physics, not policy.

Try this now:

Explain hooks like I'm not a programmer, then suggest two hooks that
would make sense for my workflow, based on what you know about how
I work. If I like one, set it up for me.

---

Part 4: Safety Nets and the Meter

19. Git in plain words

Plain English: Git is a system that snapshots your folder over time, so you can see every past version and restore any of them.

Developers guard git jealously, but it's really just superpowered version history for a folder. Every time you (or Claude) "commit," git photographs the entire folder's state. The photo album is permanent: you can flip back to any snapshot, see exactly what changed between any two, and restore anything.

If you've used "version history" in Google Docs, you already understand git — except it covers a whole folder of files at once, and snapshots happen when you choose, with a note attached ("added Q3 pricing," "rewrote intro"). You don't need to learn git commands. You need one sentence: "Commit this with a note about what we changed." Claude handles the mechanics.

Business example: A content team keeps their entire website copy in a git-tracked folder. Every revision is a commit. When a stakeholder asked "what did the pricing page say in March?" the answer took thirty seconds, not an archaeology expedition through files named final_v3_REAL_final.docx.

Try this now:

Is this folder tracked with git? If not, set it up and explain what
just happened in plain English. Then commit the current state as our
starting snapshot with a clear note.

20. Checkpoints (the undo button for AI work)

Plain English: Checkpoints are automatic snapshots taken as Claude works, so you can rewind if you don't like where things went.

This is the concept that converts nervous beginners into confident users. As Claude edits your files, Claude Code quietly saves restore points. Ask for a rewrite, hate the result? Rewind to the checkpoint before it. It's the same emotional safety as autosave in a video game: you can attempt the boss fight — the ambitious restructure, the aggressive edit — because failure costs a reload, not your progress.

Checkpoints and git overlap on purpose. Checkpoints are the automatic short-term safety net during a session; git commits are deliberate long-term milestones you choose. Belt, suspenders.

Business example: A founder let Claude take three genuinely different swings at her homepage copy — bold, conservative, playful — rewinding between attempts and keeping the winner. Without checkpoints she'd have made timid, incremental edits instead of exploring.

Try this now:

Before you rewrite this document, confirm we can rewind if I hate
the result. Then give me your boldest version — I want to see what
a big swing looks like, knowing we can undo it.

21. Tokens

Plain English: Tokens are the units AI text is measured in — roughly three-quarters of a word each. They're how usage is counted, like kilowatt-hours for electricity.

Everything Claude reads and writes is chopped into tokens (a 1,000-word document is roughly 1,300–1,400 of them). Tokens matter for exactly two reasons. First, the context window — Claude's whiteboard — is measured in tokens, so token bloat is why long sessions get forgetful. Second, usage is metered in tokens, so token bloat is also where cost comes from.

The practical insight most people miss: it's not your typing that consumes tokens — it's what Claude reads. Pointing it at a 200-page PDF to answer one question is like running the oven all day to make toast. Tell it which chapter. Give it the summary instead of the archive. Precision is efficiency; I've collected my full bag of tricks in how to cut Claude token usage in half.

Business example: A researcher stopped saying "look through all my files and find mentions of X" and started saying "check the three interview transcripts from June." Same answers, small fraction of the tokens, faster too.

Try this now:

Roughly how many tokens has our session used so far, and what
consumed most of them? Give me three habits that would make my
sessions cheaper without making the results worse.

22. Cost and plans (what this actually costs)

Plain English: You either pay a flat monthly subscription with generous usage limits, or pay per token like a utility bill. Most business users should start with the subscription.

Two ways to pay. A subscription (the Pro or Max plans) gives you a monthly price and a usage allowance that refreshes — predictable, budgetable, and hard to accidentally overspend. Pay-as-you-go API billing charges per token — scales beautifully, but the meter is always running. For a normal professional, my advice is boring and firm: subscription first. Learn your usage patterns on a fixed bill.

The framing that matters more than the price tag: compare cost to time saved, not to zero. If a subscription costs less than one hour of your billable rate per month and saves you ten hours, the spreadsheet has spoken. Use /cost inside Claude Code to watch usage, and model choice (next concept) to control it.

Business example: A copywriter tracked one month honestly: subscription cost versus 23 hours saved on first drafts and research. She raised her client capacity by a third without hiring.

Try this now:

Based on how I've been using you, estimate my usage pattern. Am I
a light, medium, or heavy user? What habits would keep me
comfortably inside a subscription plan's limits?

23. Models: Opus, Sonnet, and Haiku

Plain English: Claude comes in three sizes. Opus is the deep thinker, Sonnet the strong all-rounder, Haiku the fast lightweight. You can switch anytime with /model.

Think of staffing a project. Opus is the senior partner: best judgment, best on genuinely hard problems — strategy, subtle analysis, high-stakes writing — and the most expensive hour in the building. Sonnet is the excellent senior associate: fast, sharp, right choice for ~90% of real work. Haiku is the quick junior: ideal for simple, high-volume tasks — sorting, extracting, reformatting, summarizing at scale — at a fraction of the cost.

The workflow habit that separates pros from tourists: match the model to the task, and don't send the senior partner to do data entry. Sonnet as your daily driver; Opus when you're stuck or the stakes are high; Haiku for bulk chores. (Subagents make this concrete: a Sonnet team lead dispatching Haiku staff for grunt work is a beautiful, cheap machine.)

Model	Think of it as	Best for	Cost
Opus	Senior partner	Strategy, hard problems, high-stakes writing	Highest
Sonnet	Senior associate	Everyday real work — the default	Moderate
Haiku	Fast junior	Bulk sorting, extraction, reformatting	Lowest

Business example: A market-research workflow: Haiku subagents extract data points from 40 documents; Sonnet assembles the analysis; Opus gets one focused pass to pressure-test the conclusions. Quality where it counts, thrift everywhere else.

Try this now:

Which model are you right now? For these three tasks — [task 1],
[task 2], [task 3] — tell me which model you'd assign to each
and why, like you're staffing a small team.

---

Part 5: Automation and Honest Limits

24. Headless mode and scheduled runs (Claude working while you sleep)

Plain English: Claude Code can run without you present — triggered on a schedule or by an event, doing a predefined job and leaving the results for you.

Everything so far has been conversational: you talk, it works. But once a workflow is polished — instructions in CLAUDE.md, method in a skill, permissions dialed in — you can take yourself out of the loop. A "headless" run is Claude executing a standing order with nobody at the keyboard: every morning at 7, read the new files in the reports folder, produce the summary, save it to today's date.

This is the top of the staircase this whole article has been climbing: chat → repeatable workflow → automation. You climb it one polished workflow at a time, and only after a workflow has proven itself under supervision. Don't automate anything you haven't watched succeed ten times.

Business example: A small e-commerce owner has a scheduled morning run that reads the overnight orders export, flags anomalies (refund spikes, stockouts, unusually large orders), and writes a five-line briefing. She reads it with coffee. Nobody "did" it.

Try this now:

Of the workflows we've built together, which one is the best
candidate for a scheduled, unattended run? What would need to be
true before we could trust it without supervision, and what should
its standing instructions say?

25. When NOT to use Claude Code

Plain English: An honest tool description includes what it's wrong for. Here's my real list.

• Final judgment on high-stakes decisions. Legal commitments, medical questions, big financial calls. Claude is a phenomenal analyst and a wrong choice of decider. Let it prepare the briefing; you sign.
• Facts you won't verify. Language models can state wrong things fluently and confidently. For anything that will be published, sent, or relied upon: it drafts, you check. This is non-negotiable, forever.
• Data you're not allowed to share. Regulated, confidential, or covered by agreements? Resolve your compliance question before the data touches any AI tool, this one included.
• One-off trivial tasks. If it takes 30 seconds by hand and you'll never do it again, just do it. The payoff lives in repeatable work — that's the whole thesis of this site.
• Genuine relationship moments. Condolences, apologies, sensitive feedback to someone you care about. People can tell. Some words must cost you something.
• As a substitute for knowing your own business. Claude amplifies clarity and amplifies confusion with equal enthusiasm. If you can't describe what a good outcome looks like, that's thinking work — do it first, ideally with Claude as a sparring partner rather than an oracle.

Try this now:

Here are five tasks from my week: [list them]. For each, tell me
honestly: is this a good Claude Code task, a "draft only, human
verifies" task, or something I should do entirely myself? Argue
against yourself where it's close.

---

The Glossary Table

Bookmark this. It's the whole article in one table.

Term	Plain-English meaning
Claude Code	An AI assistant that can use your computer — read/write files, run tasks — not just chat
Terminal	A text window for talking to your computer; you type sentences, it responds
Desktop / web app	Friendlier packaging for the same engine — buttons and panels instead of the black window
Working folder	The folder Claude operates in — its "office" and the boundary of its world
Session	One continuous working conversation; resumable later
Context window	Claude's short-term memory — a big but finite whiteboard
/compact	Shrinks the conversation into a summary to free memory space
/clear	Wipes the conversation for a fresh start in the same folder
CLAUDE.md	Standing-instructions file Claude reads at the start of every session — the employee handbook
Memory	Notes Claude keeps about your preferences across sessions — its personal notebook
Slash command	A /shortcut that triggers a built-in function
Custom slash command	Your own saved prompt behind a /shortcut — "the usual" at your café
Skill / SKILL.md	A folder of instructions teaching Claude one specific job your way — a recipe card, loaded only when relevant
Installing a skill	Downloading someone's ready-made skill folder into your setup — hiring their playbook
MCP connector	A plug that gives Claude access to another tool (email, drive, data) — a keycard to one more room
Permissions	The approval system controlling what Claude may do without asking
Plan mode	Read-only mode: Claude proposes a full plan; nothing changes until you approve
Subagent	A helper copy of Claude dispatched to do part of a task in its own separate conversation
Hook	An automatic "whenever X, do Y" rule enforced by the system, not by Claude's memory
Git	Snapshot-based version history for a whole folder
Commit	One deliberate snapshot in git, with a note about what changed
Checkpoint	An automatic restore point saved as Claude works — the undo button
Token	The unit AI text is measured in (~¾ of a word); the basis of both memory limits and cost
Opus / Sonnet / Haiku	The three model sizes: senior partner / senior associate / fast junior
Headless run	Claude executing a predefined job on a schedule with nobody at the keyboard

---

Where to go from here

If you made it this far, you now know more about Claude Code than most people who've been using it for months — because you know the concepts, not just the buttons.

My suggested path, in order: create your first CLAUDE.md (concept 8, ten minutes, permanent payoff). Run one real task in plan mode (concept 16) to build trust. Then turn your most repetitive weekly task into a skill or custom command (concepts 11–12). That's the ladder: chat → handbook → repeatable workflow. Everything else in this article is refinement.

And when something breaks or confuses you — it will, occasionally — come back to the glossary. Ninety percent of Claude Code confusion is vocabulary confusion wearing a trench coat.

Resources & links

• /blog/claude-workflow-template — My complete, copyable template for turning any recurring task into a Claude workflow.
• /blog/six-claude-prompts-ahead-of-99-percent — The six prompts that put you ahead of 99% of users, with the reasoning behind each.
• /blog/secret-claude-code-commands — The lesser-known slash commands and shortcuts that never make the beginner guides.
• /blog/cut-claude-token-usage-in-half — Practical habits that halve your token spend without hurting output quality.
• /blog/most-popular-ai-coding-skills — A ranked tour of the most-installed skills and what each one actually does.
• /blog/spec-driven-development-ai-agents — The plan-first working method behind plan mode, applied to real projects.
• /skills/superpowers — The famous battle-tested skill collection; a one-download upgrade to how Claude works.
• /skills/karpathy-guidelines — Taste and judgment rules for AI-assisted work, packaged as an installable skill.
• /skills/context-engineering — A skill that keeps Claude's "whiteboard" clean so long sessions stay sharp.
• /skills/gsd-core — A get-stuff-done execution system for running whole projects through Claude.
• /skills/skilld — The meta-skill: it helps you build better skills of your own.

Want one practical, hype-free AI workflow in your inbox every week? Join the newsletter at alexandrrich.substack.com.