Reference

Cheat sheet

Everything worth remembering from the handbook, on one page. Print it, screenshot it, pin it above your desk. If you learn nothing else, learn this.

The one big idea

The quality of what comes out of Claude depends almost entirely on what you put in. Closing the gap between "what you asked" and "what you wanted" is almost entirely about context.

The four elements of a good prompt

Every prompt worth typing has these four things. Include all four and you are already in the top decile of prompters.

1. Context
Who, what, for whom. Without this, Claude is guessing at half the answer.
2. Example
A concrete "this is what good looks like". Worth a thousand words of description.
3. The ask
What you actually want, phrased in the shape you want it back. Not "help me with X", but "produce X in Y format for Z audience".
4. Escape hatch
Permission to flag uncertainty rather than guess. "If anything is ambiguous, ask me before filling it in." This is where most trust is won or lost.

The eight habits, in one line each

1. Ask it to ask you
Add "feel free to ask me follow-up questions before starting" to any non-trivial prompt. Highest-leverage line in the handbook.
2. Show before and after
If you have a raw input and a target format, paste both. Add "show me the first three rows so I can sanity-check".
3. Ask before you do
Before anything that will take more than ten minutes, ask Claude if it can do it, or help you do it faster. Before, not halfway through.
4. Rewrite your prompt
Type a rubbish version, ask Claude to rewrite it into a proper one, then run the rewrite.
5. Use a Project
If you have pasted the same context into Claude more than twice, it belongs in a Project.
6. Use your connectors
Slack, Drive, GitHub, Calendar. Let Claude read your real data instead of whatever you remember to paste.
7. Share what works
If something saved you an hour, post it in the team channel. "Look what Claude just did for me" is the whole format.
8. Use it outside work
Plan a holiday, debug your sourdough, draft the awkward landlord email. Intuition for work prompts is built on low-stakes use.

Three anti-patterns to notice

Suspiciously clean
Real data has edges. If every row fits neatly and every number is round, go check a handful against the source.
Cites something you didn't paste
If Claude mentions a fact, number, or source you know you didn't give it, treat it as a guess until you verify.
Vague where it matters
If one paragraph is hand-wavy and everything else is specific, that is usually where Claude did not actually know.

When the output is wrong

The fix is almost never "argue with Claude". It is one of these, in order:

1. Add context
The most common fix. Whatever context was missing, add it and rerun.
2. Ask what it needs
"What would you need to know to do this properly?" Claude will tell you.
3. Start fresh
If a long chat has drifted, a new chat with a tightened prompt often beats continuing.
4. Try a different model
Sonnet for most work, Opus for the hard writing and reasoning tasks.

Glossary, one line each

Claude
The large language model made by Anthropic. The thing you chat with.
Project
A workspace where you upload reference docs once and every chat inside already knows them.
System prompt
A standing instruction at the top of a Project or chat that shapes how Claude behaves across every message.
Context window
The amount of text Claude can hold in its head during a single conversation. Large but not infinite.
Connector
A link between Claude and another tool (Slack, Drive, GitHub, Calendar). Lets Claude see your real data.
Hallucination
When Claude states something wrong with confidence. Rare but real. Verify anything load-bearing.
Claude Code
A command-line tool that lets Claude read your code, make changes, run tests, and iterate.
Agent
Claude running a multi-step task on its own: read goal, act, observe, decide, repeat until done.

If something on this page isn't clear, the full version is in the Beginner guide and the glossary.