The fastest way to get better at prompting is to stop thinking of Claude as a thing that knows, and start thinking of it as a thing that continues. Given everything in front of it — system prompt, your message, pasted docs, prior turns — it produces the most likely helpful continuation. That's the whole job. Everything else downstream (why some prompts sing, why others flop, why CLAUDE.md matters, why grounding matters) is a consequence of that one fact.
So before we get to techniques, three things Claude isn't:
Claude is not a search engine. It doesn't retrieve from an index. It doesn't look anything up unless you wire a tool to it. Ask "which of our services has the worst p99?" with nothing attached and it won't say "I don't have data." It will give you a plausible answer. Confidently. That's not a bug — it's Claude doing exactly what it's trained to do.
Claude is not a database. It has no standing memory of your team, your codebase, your last conversation, or you. Some surfaces (Projects, Claude Code's CLAUDE.md, auto-memory in newer versions) graft persistence on top. But the base model is stateless. Default assumption: every new chat is a blank window.
Claude is not a colleague. It has no stake in being right, no pride in being wrong, no long-run model of what you care about. It won't push back on a bad framing unless you invite it to. Great news if you want a tireless collaborator; a real risk if you want a second opinion.
Anthropic's own guidance puts it more usefully: treat Claude like a brilliant new hire on their first day. They have world-class general skills. They have zero context on your project, your team, or what "good" looks like at your company. Every great prompting technique in this course is, at heart, a way to onboard that new hire faster.
The three things in every exchange
Every time you hit send, three things are in play. Keep this map in your head.
What Claude sees
System prompt + your message + pasted docs + prior turns + tool results. This is the only situational information Claude has. If it's not here, it doesn't exist.
What it was trained on
Broad world knowledge, code idioms, writing conventions, reasoning patterns. Zero company-internal anything. Frozen at training time — no live web, no live data.
What it's trying to do
Produce a response that's helpful, harmless, honest — the HHH frame from Constitutional AI. Not "be right." "Be the likely-best helpful continuation."
Constitutional AI, in 90 seconds
Claude is trained with a technique Anthropic calls Constitutional AI (CAI). In plain English: instead of humans rating every response for safety, Anthropic writes down a "constitution" — a set of principles — and trains Claude to critique and revise its own outputs against those principles. The model learns to be helpful, harmless, and honest from the principles themselves.
Three practical consequences you'll feel in your prompts:
- Claude will push back on clearly harmful asks — but it's calibrated to be helpful first. If it refuses something reasonable, the refusal is usually a misread of intent. Re-frame the ask with context and it often works.
- Claude will caveat and hedge when uncertain — unless you tell it not to. "Just answer, no hedging" is a legitimate instruction in many cases.
- Claude has a bias toward honest uncertainty — if you give it room to say "I don't know," it will take that room. Most hallucinations happen when the prompt doesn't leave that door open.
None of this is mystical. It's a training choice, and you can steer against it or with it depending on the job.
Quick check
Three thirty-second questions before we go on.
Feel it: the "unknown knowns" test
Below is a live Claude box. Ask it something specific about your team, your project, or a file you didn't paste. Watch what happens. (Spoiler: it will answer. Confidently. Probably wrong.)
This isn't Claude being broken. It's Claude doing exactly what it's designed to do — completing the most likely helpful next tokens — with nothing real to complete from. The fix is always the same:
The three failure modes, named
Most bad Claude outputs trace back to one of three failures. Once you can name them, you can fix them. Skim once; you'll reference this list forever.
Hallucination
Claude invents a plausible fact. Almost always a grounding failure — the fact wasn't in the window, so the model generated a shaped-right guess. Fix: paste the real source, or give it a tool to look it up.
Drift
Output starts strong, slowly wanders off-spec. Usually a constraint failure — the format or rules weren't strong enough to hold across a long generation. Fix: tighten constraints, ask for structured output, or break the task into steps.
Generic slop
Output is technically on-topic but reads like a LinkedIn post. A specificity failure — not enough context, not enough role, not enough concrete examples. Fix: add a few-shot example of what "good" looks like.
A tutor, if you want one
Still feels fuzzy? The topic tutor won't hand you the answer — it'll ask you one back. Three exchanges and the model clicks.