Claude Cheat Sheet
9 groups · 39 topics · Updated
Claude is a family of AI models made by Anthropic. This is an independent guide, checked against Anthropic's documentation in November 2026. Model names and prices change often, so confirm them on the official pages linked at the end.
Models
Which model to choose, and what it costs.
Current models
| Model | ID | Use it for |
|---|---|---|
| Fable 5.2 | claude-fable-5-2 |
The hardest reasoning, long agent runs |
| Opus 5.5 | claude-opus-5-5 |
Most work: the usual default |
| Sonnet 5.5 | claude-sonnet-5-5 |
Everyday coding and agents, faster |
| Haiku 5.5 | claude-haiku-5-5 |
High volume, simple tasks |
- Use the ID exactly as written, with no date on the end.
- Start with Opus 5.5. Move down for speed and cost, up for difficulty.
Prices
US dollars per million tokens.
| Model | Input | Output |
|---|---|---|
| Fable 5.2 | $8 | $40 |
| Opus 5.5 | $4 | $20 |
| Sonnet 5.5 | $2 | $10 |
| Haiku 5.5 | $0.10 | $0.50 |
- Output costs five times as much as input.
- The Haiku price is for prompts up to 100K tokens.
Limits
| Limit | Size |
|---|---|
| Context window | 1M tokens |
| Longest reply | 128K tokens |
- The context window is everything the model can see at once: instructions, history, tools and its own reply.
- Replies near the top of the range need streaming.
Tokens
- A token is a chunk of text, often part of a word.
- You pay for the tokens you send and the tokens you get back.
- A long conversation costs more each turn, because the whole history is sent again.
Prompting
How to ask so that the first answer is usable.
A good prompt
- Goal. What you want, and what it is for.
- Audience. Who will read the result.
- Context. The background a new colleague would need.
- Material. The text, data or code to work on.
- Format. Length, structure and tone of the answer.
- Done. How you will judge that it is right.
Explain the reason behind a rule. "Keep it short, it is read on a phone" works better than "keep it short".
Mark up the material
Wrap pasted material in tags so it cannot be confused with your instructions.
<report>
...paste the report here...
</report>
Summarise the report above in five
bullet points for a new team member.- Put long material first and the question last.
- Any tag name works. Be consistent.
Show an example
Write a product description in the
style of this example.
<example>
Small, sturdy and bright: a torch
that fits on a key ring.
</example>
Product: a folding umbrella.- One to three examples is enough.
- Vary them, or every answer will copy one.
System prompt or message
| Put it in | What goes there |
|---|---|
| System prompt | The role, the rules and the tone that hold for the whole conversation |
| User message | This task and its material |
Fix a weak answer
| Problem | Try |
|---|---|
| Too long | State a length: "under 100 words" |
| Too generic | Add the audience and the purpose |
| Wrong format | Show the format, or ask for JSON output |
| Makes things up | Supply the source text; allow "I don't know" |
| Misses a rule | Say it once, plainly, with the reason |
| Stops too early | Say what finished looks like |
Messages API
Calling Claude from JavaScript.
Set up
npm install @anthropic-ai/sdk
export ANTHROPIC_API_KEY="your-key"import Anthropic from "@anthropic-ai/sdk";
// reads ANTHROPIC_API_KEY
const client = new Anthropic();- Keep the key on the server. Never ship it in browser code or commit it.
First request
const msg = await client.messages.create({
model: "claude-opus-5-5",
max_tokens: 16000,
messages: [
{ role: "user", content: "Hello" },
],
});| Field | Meaning |
|---|---|
model |
Which model to use |
max_tokens |
The most the reply may contain |
messages |
The conversation so far |
All three are required.
Read the reply
for (const block of msg.content) {
if (block.type === "text") {
console.log(block.text);
}
}
msg.stop_reason; // → "end_turn"
msg.usage; // tokens in and outcontentis a list of blocks, not a string. Check each block'stype.
System prompt
const msg = await client.messages.create({
model: "claude-opus-5-5",
max_tokens: 16000,
system: "You are a patient maths tutor.",
messages: [
{ role: "user", content: "What is pi?" },
],
});A conversation
const messages = [
{ role: "user", content: "I am Asha." },
{ role: "assistant", content: "Hi Asha!" },
{ role: "user", content: "Who am I?" },
];- The API remembers nothing between requests. Send the whole history every time.
- The first message must come from
user.
Stop reasons
stop_reason |
Meaning |
|---|---|
end_turn |
Finished normally |
max_tokens |
Cut off: raise max_tokens |
stop_sequence |
Hit a stop sequence you set |
tool_use |
Wants to call a tool |
pause_turn |
Paused: send it back to continue |
refusal |
Declined for safety reasons |
Check stop_reason before trusting the reply.
Shaping the reply
Streaming, effort, structured output and images.
Streaming
const stream = client.messages.stream({
model: "claude-opus-5-5",
max_tokens: 64000,
messages,
});
stream.on("text", (text) => {
process.stdout.write(text);
});
const msg = await stream.finalMessage();- Text arrives as it is written.
- Use it for long replies: a slow non-streaming request can time out.
Effort
output_config: { effort: "low" },| Level | Use it for |
|---|---|
low |
Quick, simple tasks |
medium |
Default on Opus 5.5 and Haiku 5.5 |
high |
Default on Fable 5.2 and Sonnet 5.5 |
xhigh |
Hard coding and agent work |
max |
When being right matters more than cost |
Higher effort means more thinking, more tokens and a slower reply.
Thinking
thinking: {
type: "adaptive",
display: "summarized",
},- Current models reason before they answer and decide how much by themselves.
display: "summarized"returns a summary of that reasoning. By default it is left out.- On Fable 5.2 and Opus 5.5 thinking cannot be turned off. Lower the effort instead.
- The older
budget_tokenssetting is rejected by current models.
JSON output
import { z } from "zod";
import { zodOutputFormat }
from "@anthropic-ai/sdk/helpers/zod";
const Person = z.object({
name: z.string(),
email: z.string(),
});
const msg = await client.messages.parse({
model: "claude-opus-5-5",
max_tokens: 16000,
messages,
output_config: {
format: zodOutputFormat(Person),
},
});
msg.parsed_output.name;- The reply is guaranteed to match the schema.
parsed_outputisnullif parsing failed, so check it.
Images
content: [
{
type: "image",
source: {
type: "url",
url: "https://example.com/cat.png",
},
},
{ type: "text", text: "Describe this." },
],- Put the image before the question.
- A local file is sent as
type: "base64"with amedia_type.
Tools
Letting Claude call your code.
What a tool is
- A function you describe to Claude: its name, what it does and what input it takes.
- Claude never runs it. It asks, your code runs it, and you send back the result.
- The description is what Claude reads to decide when to use it. Make it specific.
Define a tool
const tools = [
{
name: "get_weather",
description: "Weather in a city now",
input_schema: {
type: "object",
properties: {
city: { type: "string" },
},
required: ["city"],
},
},
];Pass it as tools in the request.
The tool loop
- Send the request with
tools. - If
stop_reasonistool_use, find thetool_useblocks incontent. - Run each one with its
input. - Add Claude's reply to
messages, then ausermessage holding the results. - Send again. Repeat until
stop_reasonisend_turn.
Send a result back
messages.push({
role: "assistant",
content: msg.content,
});
messages.push({
role: "user",
content: [
{
type: "tool_result",
tool_use_id: block.id,
content: "18°C, cloudy",
},
],
});blockis thetool_useblock being answered.- Put every result for one turn in a single
usermessage. - If the tool failed, add
is_error: true. Do not leave the result out.
Built-in tools
These run on Anthropic's servers. There is no loop to write.
| Tool | type |
name |
|---|---|---|
| Web search | web_search_20260209 |
web_search |
| Web fetch | web_fetch_20260209 |
web_fetch |
| Run code | code_execution_20260521 |
code_execution |
tools: [
{
type: "web_search_20260209",
name: "web_search",
},
],Cost and speed
Paying less and waiting less.
Prompt caching
const msg = await client.messages.create({
model: "claude-opus-5-5",
max_tokens: 16000,
cache_control: { type: "ephemeral" },
system: longInstructions,
messages,
});- Caches the start of the request so the next one reuses it.
- The repeated part must be identical, character for character.
- Put what stays the same first and what changes last.
- A cached read costs about a tenth of the normal input price.
- The cache lasts 5 minutes; add
ttl: "1h"for an hour.
Check the cache
msg.usage.input_tokens;
msg.usage.cache_creation_input_tokens;
msg.usage.cache_read_input_tokens;| Field | Tokens that were |
|---|---|
input_tokens |
Not cached: full price |
cache_creation_input_tokens |
Written to the cache |
cache_read_input_tokens |
Read from the cache |
If the last one stays at 0, something near the start of the prompt changes every time, such as a date.
Count tokens first
const count =
await client.messages.countTokens({
model: "claude-opus-5-5",
messages,
});
count.input_tokens;Spend less
| Situation | Use |
|---|---|
| Same long prompt each time | Prompt caching |
| The answer can wait | Batches: half price |
| A simple task | Haiku 5.5, or lower effort |
| Replies too long | Ask for a length in the prompt |
| History keeps growing | Start a new conversation from a summary |
Errors
What went wrong, and whether to try again.
Status codes
| Code | Meaning | Retry? |
|---|---|---|
| 400 | The request is malformed | No |
| 401 | Missing or wrong API key | No |
| 402 | Billing problem | No |
| 403 | Not allowed for this key | No |
| 404 | Unknown model or address | No |
| 413 | Request too large | No |
| 429 | Too many requests | Yes |
| 500 | Problem at Anthropic | Yes |
| 529 | Service overloaded | Yes |
Catch errors
const { RateLimitError } = Anthropic;
const { APIError } = Anthropic;
try {
await client.messages.create(params);
} catch (error) {
if (error instanceof RateLimitError) {
// wait, then try again
} else if (error instanceof APIError) {
console.error(error.status);
} else {
throw error;
}
}- Check the most specific class first.
- Test the class, not the text of the message.
Retries and timeouts
const client = new Anthropic({
maxRetries: 2, // the default
timeout: 600_000, // 10 min, in ms
});- The SDK already retries 408, 409, 429, 5xx and connection errors.
- The timeout is in milliseconds.
Refusals
if (msg.stop_reason === "refusal") {
msg.stop_details.category;
msg.stop_details.explanation;
}- A refusal is a normal response, not an error. Nothing is thrown.
stop_detailsis only filled in for a refusal.
Claude Code
Claude as a coding assistant in your terminal.
Install
curl -fsSL https://claude.ai/install.sh \
| bash| System | Command |
|---|---|
| Windows PowerShell | irm https://claude.ai/install.ps1 | iex |
| Homebrew | brew install --cask claude-code |
| npm | npm install -g @anthropic-ai/claude-code |
- Check with
claude --version. - It needs a paid Claude plan or a Console account.
Start and resume
| Command | Does |
|---|---|
claude |
Start a session in this folder |
claude "query" |
Start with a first prompt |
claude -p "query" |
Answer once, then exit |
claude -c |
Continue the last conversation here |
claude -r |
Pick a past session to resume |
claude update |
Update to the latest version |
cat logs.txt | claude -p "explain"Useful flags
| Flag | Does |
|---|---|
--model |
Choose the model: opus, sonnet, haiku, fable |
--permission-mode |
Start in a mode such as plan or acceptEdits |
--add-dir |
Give access to another folder |
--output-format |
With -p: text, json or stream-json |
Commands in a session
| Command | Does |
|---|---|
/help |
List the commands |
/init |
Create a CLAUDE.md for the project |
/clear |
Start a new conversation |
/compact |
Summarise the conversation to free space |
/context |
Show what is filling the context |
/model |
Switch model |
/effort |
Set the effort level |
/plan |
Plan first, change nothing yet |
/rewind |
Go back to an earlier point |
/resume |
Reopen a past conversation |
/permissions |
Choose what runs without asking |
/mcp |
Manage connections to outside tools |
/usage |
Show cost and plan limits |
/code-review |
Review the current changes |
CLAUDE.md
- A file of instructions for the project, loaded at the start of every session.
- Put in it what a new teammate would need: how to build, how to test, the rules of the codebase.
- Create it with
/init, edit it with/memory. - Keep it short. Everything in it uses context in every session.
- How it fits a wider workflow: the AI-native SDLC.
More help
The official pages this sheet was checked against.
Official documentation
| Topic | Page |
|---|---|
| Models | Models overview |
| Prices | Pricing |
| Prompting | Prompt engineering |
| Requests | Create a message |
| Tools | Tool use |
| Caching | Prompt caching |
| Errors | Errors |
| Claude Code | CLI reference and commands |