Meet Claude: AI fluency foundations

Claude is a large language model (LLM): a system trained on enormous amounts of text to do one deceptively simple thing - predict the next word, over and over, at incredible quality. It is not a database, not a search engine, and not a mind reader. Everything impressive and everything weird about AI flows from this one mechanic.

Because it predicts rather than looks up, the same prompt can produce different answers on different runs. That variability is a feature for brainstorming and a risk for facts - which is why verification habits matter (Part 3).

Why this matters to you

• It explains fluency without guarantees: Claude always sounds confident, because confident text is the most likely next text. Confidence is not evidence.
• It explains why giving material beats asking cold: prediction conditioned on YOUR document is anchored; prediction from nothing is educated guessing.

Raw next-token prediction would happily continue any text, helpful or harmful, true or false. Models get their working personality in a second training phase: human feedback teaches the model which responses are helpful, honest, and safe. This is why Claude asks clarifying questions, declines some requests, and admits uncertainty - those behaviors were trained in deliberately.

• Different models, different characters: the same question to Claude, ChatGPT and Gemini gets different styles because each company trains different behavior, not because one "knows more".
• Character is steerable: your instructions layer on top. Tell Claude "be blunt, skip the caveats" or "challenge my assumptions" and it will - see Steerability below.
• Claude specifically is trained by Anthropic with an emphasis on being helpful, harmless, and honest - it would rather say "I'm not sure" than fabricate, but it is not immune to error. Trust calibrated, never blind.

Claude's built-in knowledge comes from training data with a cutoff date. It knows a compressed version of the public internet up to that point - and nothing after, nothing behind your company firewall, and nothing about your meeting this morning unless you paste it in.

When asked for facts beyond its knowledge, a language model's failure mode is not silence - it is hallucination: fluent, plausible, wrong. The fix is always the same: supply the source, ask it to cite what it used, and tell it that "I don't know" is an acceptable answer.

Everything in your current conversation - your messages, uploaded files, Claude's replies - lives in a working memory called the context window. It is big (hundreds of pages) but finite, and two behaviors follow:

• Long chats drift. In a marathon conversation, early details fade in influence. When a thread gets messy, start fresh and paste in a short summary of decisions so far - a clean context beats a cluttered one.
• Each conversation starts blank. Claude does not remember yesterday's chat unless your workspace has that feature enabled or you re-supply the material. Projects (Part 2) exist precisely to solve this - persistent knowledge that every conversation can see.

Steerability is the property that makes everything else usable: Claude will genuinely change its behavior - tone, format, depth, persona, rules - based on plain-language instructions. Most people dramatically under-steer. You can set rules for a whole conversation ("always answer in tables", "flag any assumption you make") and Claude will hold them.

When properties collide is where AI gets weird. Steer it hard toward confidence ("give me a definitive answer") and you suppress its honesty about uncertain knowledge - you have steered it INTO hallucination. Ask for perfect recall at the end of a very long chat and steering can't beat memory limits. Odd outputs are almost never random: one property is being pushed against another. Diagnose which, fix the prompt.

A Claude conversation is a working session, not a query box. The habits that change everything:

• Talk to it like a bright new colleague - full sentences, real context, actual stakes. "Help me prep for a tense budget meeting with marketing tomorrow" beats "budget meeting tips".
• Attach, don't describe. Drop in the file (PDF, spreadsheet, screenshot, code) instead of describing it. Claude reads images and documents directly.
• Iterate in the same thread. "Shorter." "More formal." "Now as a table." Each follow-up refines - you are steering, not re-rolling dice.
• Edit your prompt when a thread goes sideways - rewriting your message and re-running beats arguing with a bad draft.
• Pick the right model: the default model balances speed and depth; switch to the most capable model for complex reasoning, the fast one for bulk simple tasks.

A Project is a workspace with permanent memory: upload your team's key documents once, set standing instructions once, and every conversation inside the Project starts already knowing them. This is the single highest-leverage feature for teams - it turns Claude from a stranger into a briefed colleague.

What goes in a good project

• Knowledge files: style guides, product specs, org charts, glossaries, key reports - the stuff you're tired of re-explaining.
• Custom instructions: "You support the data science team at [company]. Our stack is X. Always flag assumptions. Use our metric definitions from the glossary."
• Team sharing: on the Teams plan, share the Project so everyone's conversations start equally briefed - one setup, whole-team payoff.

Artifacts turn Claude's output into a live, standalone piece of work beside the chat: documents, tables, diagrams, working web pages, small interactive tools. Ask for "an interactive checklist my team can tick off" and you get a working mini-app, not a paragraph. Artifacts can be edited by conversation ("make the header plum, add a filter"), shared with a link, and are how we prototype websites in Session 4.

Skills are reusable instruction packs that teach Claude a repeatable job - your report format, your QA checklist, your SOP - so you invoke them instead of re-explaining. They are the star of Session 3, where we turn one of OUR real SOPs into a skill live.

• Connectors plug Claude into your tools - calendar, email, drive, project trackers - so it can read and act on live data. On our Teams plan most connectors need a one-time sys-admin approval; Session 2 includes the ready-to-send request template and builds the flagship automation (calendar → prioritized to-do list).
• Web search closes the knowledge-cutoff gap: Claude looks up current information and cites sources. Use it whenever "recent" matters.
• Research mode goes further: Claude plans a multi-step investigation, reads many sources, and returns a structured, cited report. Right for "compare these 5 vendors" - overkill for "what does churn mean".
• Enterprise search (when enabled by the org) lets Claude answer from internal company sources - ask your admin what's connected before assuming.

Beyond the browser: the desktop app adds Cowork (Claude works on files in a folder you choose) and hosts Claude Code for engineers; the mobile app handles voice conversations and photos of whiteboards; Claude also lives inside some tools your team already uses. Session 5 covers Claude Code properly - here's what each seat in this room reaches for first:

Why "fluency" and not "prompting"? Because prompting is only one D of four. Fluent people also choose the right work to hand over, judge outputs like an editor, and stay accountable for what ships. The four Ds, in working order:

Delegation has three layers, and most people skip the first two:

• Problem awareness: what are you actually trying to achieve, and what does "good" look like? Fuzzy goals produce fuzzy delegations.
• Platform awareness: what is AI genuinely good at (Part 1!) and which tool fits - chat, a Project, research mode, or a human colleague?
• Task delegation: split the work. Which parts go to Claude, which stay with you, which are collaborative ping-pong?

Project planning with delegation in mind

For any real project, list the tasks and mark each: AI-first (drafting, summarizing, structuring, variations), human-first (final judgment, relationships, confidential calls), or duet (AI drafts, you shape; you outline, AI expands). A campaign launch might delegate audience research summaries and copy variants, keep budget sign-off and partner negotiations, and duet the creative brief.

Description quality is the single biggest lever you control. Three kinds of describing, then the techniques:

• Product description: what you want - task, format, length, audience.
• Process description: how to work - "think step by step", "ask clarifying questions first", "give 3 options before committing".
• Performance description: how to behave - tone, persona, confidence level, what to do when unsure.

The six elements of a strong prompt

Techniques that punch above their weight

• Show an example of what good looks like (last week's report) - one example outperforms three paragraphs of description.
• Ask it to ask: "ask me up to 3 questions before you start" catches missing context early.
• Positive instructions: say what TO do ("plain, confident language") over what not to do ("don't be wordy").

Discernment is editorial judgment over three things: the product (is it right, complete, on-brand?), the process (did it reason soundly or leap?), and the performance (was the interaction style right - too agreeable, too verbose?).

The core working rhythm of AI fluency is the description-discernment loop: describe → get output → judge → describe better → repeat. First drafts are the start of the conversation, not the verdict. Specific feedback ("the tone in section 2 is alarmist - make it neutral and factual") beats re-rolling ("try again") every time.

Diligence is the D that protects your name and the company. Three parts:

• Creation diligence: be thoughtful about which AI you use for what, and what data you share with it. Company policy and client confidentiality come first - when in doubt, sanitize.
• Transparency diligence: be appropriately open about AI's role in your work - with colleagues, clients, and regulators where it matters. "Claude drafted, I verified and decided" is a sentence to be comfortable saying.
• Deployment diligence: verify before it ships. You own the output - the Avianca lawyers learned this in federal court.

Claude adapts to your level, never judges a "dumb" question, and can quiz you until a concept sticks. The trap to avoid: letting AI do the understanding FOR you. Copy-pasting an explanation you didn't absorb is the professional version of copying homework - the goal is that YOU get smarter, not that an answer got produced.

Four techniques, in rising order of learning power

• Explain: "Explain X like I'm a marketing director, not an engineer."
• Socratic: "Don't give me the answer - ask me questions until I get there."
• Quiz: "Quiz me with 5 questions, one at a time, and grade my answers honestly."
• Reverse: "I'll explain it back to you - correct my misunderstandings." (The strongest: teaching reveals what you don't know.)

Career planning, the fluent way

The same techniques apply to your own growth: "Here's my role and where I want to be in 2 years - interview me about my skills, then give me a gap analysis and a 90-day learning plan using free resources." Claude is a tireless, honest career-planning partner - the decisions stay yours.

The through-line of everything today: AI changes what you can produce, not what you're responsible for. The judgment calls - what to work on, what's true enough to ship, what's right for the client, what your name stands behind - remain human work, and becoming fluent makes your judgment MORE valuable, not less. The people who thrive with AI aren't the ones who automate the most; they're the ones who know exactly which 20% must stay human.

If you lead people, you are now an AI educator whether you planned it or not. The teaching course's core insight: the 4Ds work as two loops, and people learn them by doing, not by hearing.

• The delegation-diligence loop (the strategic loop): what should we hand to AI, and can we stand behind the result? Run it as a team exercise on YOUR team's task list - 30 minutes, whiteboard, three columns: AI-first / human-first / duet.
• The description-discernment loop (the craft loop): prompt, judge, re-prompt. Teach it by pairing: one person drives, the other critiques the output, swap. The critique seat learns fastest.
• Assess fluency by artifacts, not vibes: ask each person to bring one before/after (their old way vs their 4D-prompted way) to a team meeting. The delta is the assessment - and the demos become the team's shared playbook.
• Apply your discipline's expertise: every field has quality standards AI doesn't know by default (your data QA bar, your brand voice, your legal review). Fluency in YOUR team means encoding those standards into prompts, Projects, and later skills (Session 3).