I Passed Anthropic's Official Certifications. Here's What 90% of Claude Users Don't Know.

Introduction

Everyone uses Claude. Very few people use it well.

Several official certifications on Anthropic Academy — Anthropic’s free learning platform launched in March 2026 — were completed. These courses were built by Anthropic’s internal teams: no reinterpretation, no fluff, no oversimplification.

What was learned forced a complete rethink of how to interact with Claude.

1. The Real Problem: You’re Automating When You Should Be Augmenting

Before discussing features or prompts, Anthropic draws a fundamental distinction:

• Automation = AI does the work for you. You check out.
• Augmentation = You and AI collaborate as thinking partners. You stay in the loop.
• Agency = You configure AI to act autonomously within boundaries you have defined.

Most users are stuck in Automation mode — they type an instruction, copy-paste the result, and move on. The output is generic. The dependency grows. And they cannot tell if Claude did a good job or a bad one.

Anthropic Academy calls this the “Human in the Loop” principle: you decide what to ask, you evaluate the outputs critically, you bring judgment, creativity, and ethics. Claude amplifies your capabilities — it does not replace them.

Do not treat Claude like a vending machine. Treat it like a smart colleague. You still have to think.

2. Why Claude Behaves the Way It Does (The Official Explanation)

Anthropic published this in their official AI Capabilities & Limitations framework. Four core properties explain almost every behavior from Claude — both impressive and frustrating.

Property 1: Claude Is Not Looking Up Answers. It’s Writing Them.

At its core, Claude does one thing: given everything written so far, it predicts what comes next — one word at a time, billions of times, shaped by everything it learned during training. It is closer to a vastly sophisticated autocomplete than a search engine.

When the task is common and well-represented in its training data (summarize, explain, reformat), it works beautifully. When the task requires distinguishing a true fact from a plausible-sounding one — especially in niche or technical domains — it can generate confident nonsense.

Hallucinations happen because Claude is not lying. It is completing a pattern. The pattern just happens to be wrong.

Simple version: Claude is writing an answer, not finding one.

Property 2: Claude’s Knowledge Has Edges — And They’re Predictable

Claude learned by processing enormous amounts of text. That knowledge was frozen at a specific moment: the knowledge cutoff date. But Claude’s knowledge is uneven in a specific, predictable way.

Topics that appeared frequently, recently, and consistently in training data → Claude knows well. Topics that are rare, recent, niche, or contested → Claude drifts toward the edge. This is where it is most likely to sound confident and be wrong.

The right question is not “does Claude know this?” but “how well-represented was this topic in what it read?”

Verify anything recent, niche, or where being wrong has real consequences. Do not trust Claude’s confidence — verify the claim independently.

Property 3: Claude Forgets Between Sessions (And Has a Hard Limit)

Everything Claude can work with sits inside a fixed-size workspace called the context window — instructions, uploaded documents, its own previous responses, the conversation history. When that window fills up, something falls off. By default, when you start a new conversation, the window resets completely. Claude has no memory of yesterday unless you explicitly give it that memory (via Projects, covered below).

This property does not degrade gradually. It is a cliff. Things work until they suddenly do not.

Practical implications: Very long conversations cause loss of early context. Very long documents have parts dropped silently. New sessions are a complete reset unless you use Projects.

Property 4: Claude Follows Instructions — But Not the Way You Think

Claude is remarkably steerable. You can give it a role, a format, a tone, and constraints — and it applies them. But steerability is not the same as understanding. Claude follows your instructions the same way it does everything else: by continuing a pattern. There is always a gap between what you intended to communicate and what actually landed.

Control is highest when instructions are short, concrete, and verifiable: “respond as a table,” “keep it under 100 words,” “use exactly this format.” Control degrades when instructions get abstract, when the chain of reasoning gets long, or when the task requires real precision (math, formal logic) that pattern-matching does not reliably supply.

The question is not “did I give good instructions?” but “how much room is there between my words and my actual intent?”

How the 4 Properties Interact

Most real failures are two properties colliding:

• Hallucinated citation = Next Token Prediction + Knowledge gap
• Confidently wrong math = Next Token Prediction + Steerability
• Drift in a long conversation = Working Memory fading + Steerability
• Claude agreeing with a bad premise = Trained sycophancy + Next Token Prediction

You do not need to memorize every failure mode. A small mental model is enough. When something goes wrong, you can recognize which kind of wrong it is and respond accordingly.

3. The 4D Framework: The Real Grammar for Working with Claude

Anthropic co-developed with academic researchers a complete framework called AI Fluency. At its core: the 4Ds. These are not four prompting tips. It is a mental architecture for any serious AI collaboration.

D1 – Delegation: Think Before You Type

Delegation means deciding who does what before you touch the keyboard.

Three levels:

• Problem Awareness – Clarify your actual goal before involving AI. What is the task? What does success look like?
• Platform Awareness – Understand what the tool can and cannot do. Claude is not Google, not a search engine, not an oracle. It has a knowledge cutoff. It can hallucinate.
• Task Delegation – Split the work intelligently. Anything requiring human judgment, ethical decisions, or deep personal context stays with you. Anything requiring speed, consistency, or large-scale information processing goes to Claude.

Use it when: Starting any new task with Claude. Ask “what exactly do I need here, and what part of this should I actually be doing myself?”

Do not use it when: You already have a clear, simple, one-shot task like “translate this sentence” or “fix this typo.” Overthinking delegation for tiny tasks wastes time.

Concrete example: You want Claude to write a job offer email.

• Bad delegation: “Write me a job offer email.”
• Good delegation: “Draft the structure and tone. I will fill in the actual salary and personal details, because only I know what we agreed on.”

D2 – Description: It’s Not a Prompt. It’s a Conversation.

Description is how you communicate with Claude. It is not “prompting” in a technical sense — it is a collaborative conversation.

Three dimensions:

• Product Description – Tell Claude exactly what you want in the output: length, audience, tone, format, style.
• Process Description – Guide how Claude should approach the problem: “think step by step,” “consider multiple angles before concluding,” “start with counterarguments, then build the case.”
• Performance Description – Define Claude’s behavior during the interaction: “be a critical editor,” “play the role of a skeptical investor,” “ask me questions instead of answering directly.”

Use it when: You need a specific, high-quality output. The more complex the task, the more all three dimensions matter.

Do not use it when: Simple factual questions. “What is the capital of France?” does not need a performance description.

Concrete example:

• Before: “Summarize this article.”
• After: “Summarize this article in 5 bullet points. Audience: non-technical executives. Focus on business impact, not technical details. Flag anything that seems uncertain or that I should verify.”

D3 – Discernment: Evaluate. Don’t Just Accept.

Discernment is your ability to critically evaluate what Claude produces and use that evaluation to improve the next iteration.

Three levels:

• Product Discernment – Quality of what Claude produced. Is it accurate? Is it relevant? If Claude sounds confident, that does not mean it is right. Verify.
• Process Discernment – How Claude reasoned. Are there logical errors? Hidden assumptions? Ask Claude: “Walk me through your reasoning on this.”
• Performance Discernment – Claude’s behavior during the exchange. Did it adopt the right role? Was it too agreeable? Too vague?

Description + Discernment = the feedback loop that transforms Claude from a tool into a thinking partner.

Use it when: Always. Every single output. Even if it looks good at first glance.

Common mistake: Reading Claude’s response once, thinking “looks fine,” publishing it. That is Automation mode, not Augmentation.

Concrete example: Claude gives you a security recommendation. Before applying it, ask: “What assumptions are you making here? What would change if the system is air-gapped?” That is Process Discernment.

D4 – Diligence: You Own Everything You Publish

Diligence is the responsibility you keep over everything you publish, share, or deploy with Claude’s help.

Three aspects:

• Creation Diligence – Choose the right AI tools based on confidentiality, security, and context. Do not paste client data, credentials, or proprietary code into a tool you have not vetted.
• Transparency Diligence – Be honest about AI’s role in your work with those who need to know. “AI-assisted” is different from “AI-generated.” Know the difference and communicate it.
• Deployment Diligence – Take ownership of outputs you use or share. Verify accuracy. Ensure appropriateness. Never publish something you cannot defend because “Claude wrote it.”

Use it when: Before sharing, publishing, or deploying anything Claude helped produce.

The trap: Treating diligence as optional because “everyone uses AI now.” Your name is on it. You are responsible.

4. The Features 90% of Users Never Touch

Projects: Your Persistent Memory

A Project in Claude is a space with shared, persistent memory across all your conversations. You store context once — your role, style, constraints, reference files — and Claude accesses it automatically every time.

Without Projects, you re-explain your context every new conversation. With Projects, you always start at the right level.

How to use it: Create one Project per domain (work, personal project, research). Add a custom system prompt with your role, preferences, and recurring context. Every conversation inside that Project inherits it automatically.

When to use it: Any recurring workflow — client work, ongoing research, a codebase you keep coming back to.

When not to use it: One-off questions that do not need any particular context.

Artifacts: Deliverables, Not Just Text

When Claude generates an artifact — a document, a table, code, a structured analysis — it is an object you can edit, iterate on, export, and reuse directly. The artifact is live-editable. You can ask Claude to modify one specific section without regenerating everything.

How to use it: After Claude generates something, click into the artifact and ask for targeted edits: “Change only the conclusion section” or “Add a column for risk level to this table.”

When to use it: Any output you will actually use somewhere — reports, code, templates, structured analyses.

When not to use it: Conversational exchanges. Not every answer needs to be an artifact.

Skills: Automate Your Own Workflows

Skills are reusable Markdown instructions that Claude applies automatically when you invoke them. You create them once, reuse them indefinitely.

Example: a “security code review” Skill that systematically checks inputs, attack surfaces, and dependencies. Invoke it with one word, Claude runs the full protocol.

How to use it: Write your recurring instructions in a Markdown file, save it as a Skill, give it a name. Next time you need that workflow, just call it.

When to use it: Any task you find yourself explaining to Claude more than twice. If you are copy-pasting the same instructions repeatedly, that is a Skill waiting to be created.

When not to use it: One-time tasks that will never repeat.

Research Mode: Actually Going Deep

Research Mode is not a web search. It is a deep investigation mode where Claude runs dozens of iterative searches, synthesizes sources, identifies contradictions, and produces a structured report with citations.

When to use it: Market analysis, technical state-of-the-art reviews, competitive intelligence, due diligence on a technology or vendor.

When not to use it: Simple factual questions. “What year was Python created?” is overkill.

Connectors & MCP: Claude Connected to Your Real World

Connectors let Claude access your tools directly — Google Drive, Slack, GitHub — without copy-pasting. The Model Context Protocol (MCP) is Anthropic’s open standard for connecting Claude to any external system via three primitives: tools (actions Claude can take), resources (data Claude can read), and prompts (reusable instruction templates).

MCP turns Claude from a text assistant into an agent that can actually act inside your ecosystem.

When to use it: When you want Claude to work with your existing tools, not alongside them.

When not to use it: If you are just chatting or doing one-off tasks, Connectors add unnecessary complexity. Start simple.

5. Claude Code: It’s Not a Copilot. It’s an Agent.

Most developers use Claude Code like an upgraded Copilot — ask it to write code, copy it, move on. That misses the point entirely.

Claude Code works through a tool system: it reads files, runs commands, modifies code, executes tests — sequentially and with reasoning. This is not text generation. It is action inside your environment.

CLAUDE.md – Your project’s memory

A configuration file that defines the project context for Claude Code. Three levels:

• Project level: shared with the team, committed to source control
• Local level: personal preferences, not committed
• Machine level: global instructions across all your projects

Run /init on first launch. Claude analyzes your codebase and creates the file automatically. Then customize it.

When to use it: Every project. Always. This is the single biggest quality-of-life improvement in Claude Code.

Plan Mode (Shift + Tab twice) – Think before acting

Forces Claude to deeply analyze the task and create a detailed implementation plan before taking any action.

When to use it: Multi-step tasks, anything touching multiple files, refactors, or features you do not fully understand yet.

When not to use it: Simple, isolated changes where you just want the edit done.

Thinking Mode (“Ultra think”) – Going deep on hard problems

Gives Claude an extended reasoning budget for complex logical problems. Plan Mode = breadth (understanding the whole codebase). Thinking Mode = depth (solving a tricky specific problem). They can be combined.

When to use it: Tricky debugging, complex algorithms, architectural decisions.

When not to use it: It costs more tokens. Do not use it for trivial tasks.

Custom Commands – Your personal slash commands

Create a file in .claude/commands/. An audit.md file automatically creates the /audit command. You write the instructions in Markdown, Claude executes the full workflow when you call it.

When to use it: Any repetitive workflow — dependency auditing, test generation, security checks, PR descriptions.

Hooks – Automated feedback loops

Scripts that run before or after every tool call Claude makes. A post-edit hook that runs tsc --no-emit after every TypeScript file change, feeds the errors back to Claude, and forces automatic correction of all call sites. Another hook spins up a separate Claude instance to detect duplicate code. This is self-correction on a loop.

When to use it: Large projects where Claude tends to create type errors or duplicate code. Critical directories where quality matters most.

When not to use it: Small projects or early exploration. Hooks add overhead — only apply them where the trade-off makes sense.

GitHub Actions Integration

Claude Code integrates directly into your pull requests via /install-github-app. It can automatically review PRs, flag security risks (PII exposure, vulnerabilities), and respond to issues when mentioned with @Claude.

Claude Code SDK – For pipelines and automation

A programmatic interface (CLI, TypeScript, Python) to integrate Claude Code into your existing pipelines. Default permissions are read-only — write permissions must be explicitly enabled.

When to use it: Building hooks, helper scripts, or integrating Claude Code intelligence into larger automated workflows.

6. What This Actually Changes

The real lesson from these certifications is not a list of features. It is a shift in posture.

Most people ask a question, read the answer, and move on. The 4D framework teaches you to build rich context, guide the reasoning, critically evaluate the output, and iterate intelligently.

Claude is not a search engine. Claude is not a text generator. Claude is a cognitive partner — but only if you know how to collaborate with it.

The difference between someone who knows these frameworks and someone who does not is not the tool. It is the mental model.

Go Further

All Anthropic Academy courses are free, self-paced, with an official certificate on completion.

Courses completed:

• Claude 101 – the universal starting point
• Claude Code in Action – for developers
• AI Fluency: Framework & Foundations – the full 4D framework
• Introduction to Agent Skills – Skills & automation
• Introduction to Subagents – multi-agent architectures
• AI Capabilities and Limitations – understanding what Claude can and cannot do

Full platform: anthropic.skilljar.com