Anthropic’s prompt engineering guide for 2026 opens with a short list of core techniques, then does something less common: it names two techniques most tutorials still lead with, and explains why they matter less than they did.
Prompting the newest models is, in the guide’s own phrase, “less scaffolding, more curation”.
What the guide says the goal is
“the best prompt isn’t the longest or most complex. It’s the one that achieves your goals reliably with the minimum necessary structure.”
That is a reversal. Prompt engineering was taught as an additive craft: add a role, add tags, add a step-by-step instruction, add examples. The guide keeps what changes output and demotes what mostly adds tokens. The Claude getting-started guide covers the product side.
Core techniques that pay off first
Be explicit about the output, not the topic. A dashboard request becomes one asking for as many relevant features and interactions as possible, plus a line saying to go beyond the basics.
Give the motivation too. “NEVER use bullet points” is a rule with no reasoning attached; the replacement explains that the reader finds flowing prose easier to read. A rule the model understands generalises to cases you did not anticipate. Give explicit permission to say “I don’t know” as well, which the guide ties to fewer confident errors.
How many examples to give
Examples are, in the guide’s words, one of the most reliable ways to steer format, tone and structure.
“Start with one example (one-shot). Only add more examples (few-shot) if the output still doesn’t match your needs.”
One, not five, is the default. Current models also pay close attention to details in examples, which cuts both ways: a careless example teaches the wrong pattern as well as a good one teaches the right one.
What a well-built prompt contains
Long prompts get unreadable for the same reason long functions do: everything sits at one level.
The four blocks
Instructions: what to do, in the imperative. Context: why it matters and who the reader is. Examples: one or two, wrapped so they cannot be mistaken for instructions. Input: the real material, marked clearly.
Ordering advice is precise. With long inputs, put the material above your query and keep critical details at the beginning or the end. XML tags were the old way to draw those boundaries; headings, whitespace and explicit wording now do the job with less overhead.
Advanced techniques, and one to skip
Prefilling is the most underused technique on the list, chaining the one that costs and pays the most, and role prompting the one to handle carefully. A helpful assistant often beats an expert who only speaks in jargon.
Prefill, which most people miss
Prefilling means starting the model’s answer for it. In a chat box you ask for valid JSON with no preamble and tell the model to begin with an opening brace. In the API you do it literally, by ending the message list with an assistant turn.
import anthropic
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=512,
system="Extract figures from the report. Output JSON only, with no commentary.",
messages=[
{"role": "user", "content": report_text},
{"role": "assistant", "content": "{"},
],
)
print(message.content[0].text)The assistant turn is the prefill: the model continues from that brace instead of deciding how to open. Chain of thought is the technique the guide most clearly updates, no longer default advice because extended thinking is preferred where the model supports it. Chaining is the one technique that cannot fit in a single prompt; it splits a task into sequential calls and pays extra latency for better accuracy.
Worked example: one prompt, rebuilt
Take a request most people write in one line, and watch what each pass adds.
Summarise the attached quarterly report.The baseline is not wrong, it is underspecified: no audience, no length, no format, and nothing about a missing figure.
Summarise the attached quarterly report for an investment committee.
Keep it under 300 words, in flowing prose, with no bullet lists.
If any figure you need is not stated in the report, write "not stated" instead of estimating.Pass two adds the audience, the constraint, the format and the honesty clause. Three of those are core techniques; the honesty clause is permission to be uncertain, which the guide links to fewer invented answers. Pass three is chaining: extract the figures first, then write the summary from that extraction. A summary built on checked numbers is a different product from one built on a guess.
Where these prompts run in code
Two techniques are easiest to see in a request body: the role, which belongs in the system prompt, and chaining, which is two calls.
import anthropic
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
system="You are a helpful coding assistant specializing in Python.",
messages=[
{"role": "user", "content": "How do I sort a list of dictionaries by key?"}
],
)
print(message.content[0].text)That system prompt sets a focus in one sentence rather than stacking adjectives, because the warning about over-constrained roles applies exactly here.
import anthropic
client = anthropic.Anthropic()
extract = client.messages.create(
model="claude-sonnet-5",
max_tokens=512,
messages=[
{"role": "user", "content": f"List every figure in this report.\n\n{report_text}"}
],
)
summary = client.messages.create(
model="claude-sonnet-5",
max_tokens=512,
system="Summarise for an investment committee in under 300 words.",
messages=[{"role": "user", "content": extract.content[0].text}],
)
print(summary.content[0].text)The model id is configuration, and the published line-up with per-token prices is on the plan comparison page. Field names follow the API examples shipped with the official SDKs; the API reference did not open from the machine used here, so confirm anything unfamiliar against it.
When a prompt fails, and the fix
A generic response usually means the prompt asked for something generic. A response that misses the point means the goal stayed in your head. An inconsistent format is an examples or prefill problem, and an unreliable result on a complex task is a signal to chain rather than to add more instructions.
Unwanted preambles are fixed by asking for the answer directly, and invented details by saying that “not stated” is acceptable. Longer prompts are also billed, and prompt caching only helps a stable prefix, which the API getting-started guide covers.
Frequently asked questions
Do these techniques apply to Claude Code too? They are about the model, so they carry over. Instructions for every session belong in a CLAUDE.md file, and the terminal guide says where it goes.
Do I still need XML tags? Not by default. The guide keeps them for genuinely ambiguous boundaries.
Is chain of thought obsolete? Demoted, not obsolete. Extended thinking is preferred where the model supports it.
How many examples is right? One, until one is not enough.
Cut the prompt in half, then add back only the blocks that change the answer. That is the 2026 position in one sentence.