Mastering Claude (Anthropic) · Claude foundations · lesson 2 of 20 · 14 min

Chat fundamentals: the prompt structure experts use

Why structure beats cleverness

There are no magic words that unlock Claude. What reliably improves results is giving the model the same things you would give a talented new colleague: who the work is for, what you want, the rules, and what "done" looks like. Anthropic's own prompting guidance emphasises being clear and direct, giving context and motivation, using examples, and separating materials with XML-style tags. This lesson turns that into a habit you can apply in thirty seconds.

The C-T-C-F structure

For any prompt that matters, include four parts:

Context: who you are, who the audience is, the situation and why the task matters. ("This goes to procurement managers who skim on mobile.")
Task: one clear job, stated as an instruction. ("Draft a follow-up email that confirms the three agreed next steps.")
Constraints: length, tone, spelling, things to avoid, facts that must not change, compliance rules.
Format: the shape of the output: a table with named columns, H3 headings, a 5-bullet summary, JSON, an email with a subject line.

Claude's newer models follow instructions very precisely, so say what you want rather than hinting. If you want suggestions only, say "suggest, don't rewrite". If you want it to make changes, say "rewrite".

Tags keep materials separate

When you paste several things into one message, wrap each in a labelled tag. Claude is trained to pay attention to this structure.

<brief>
Launch of our Ramadan gift boxes, UAE and KSA, 3 price points.
</brief>

<brand_rules>
Warm, respectful, no discount language before the first week. Arabic and English versions.
</brand_rules>

<past_posts>
[paste 3 approved posts]
</past_posts>

Using the brief and brand rules, write 5 Instagram captions in the style of the past posts.
Return a table: Caption (EN) | Caption (AR) | CTA | Why it fits the brand rules.

Tags also let you refer to materials precisely later ("check each caption against <brand_rules>").

Examples are the strongest steering tool

One or two short examples of the output you want (sometimes called few-shot prompting) beat paragraphs of description. Label them clearly and say what to copy (structure, tone, length) and what not to copy (the specific facts).

Let Claude ask questions

End important prompts with: "Before you start, ask me up to three questions if anything important is missing." This turns guessing into a short interview. It is the single cheapest quality boost there is.

Iterate like an editor, not a slot machine

When a draft is close, give specific feedback: "Keep the structure. Cut to 150 words. Remove the second paragraph. Make the CTA a question." If the draft is wrong because your prompt was missing something, edit your original message and resend rather than stacking corrections; a clean context produces a cleaner answer. Use "retry" only when the prompt was good and you want a different alternative.

Worked example: from vague to production-ready

Vague: "Write a LinkedIn post about our new service."

Structured:

<context>
I run a 6-person bookkeeping firm in Manchester serving e-commerce sellers.
We have launched a monthly VAT-readiness check. Audience: UK Shopify and Amazon
sellers turning over £90k–£500k who worry about VAT registration thresholds.
</context>

<task>
Write a LinkedIn post announcing the service.
</task>

<constraints>
- 120–160 words, British spelling, plain English, no hype words ("game-changer", "revolutionary").
- Do not state tax thresholds or rates; say "check the current HMRC threshold" instead.
- One question to invite comments.
</constraints>

<format>
Hook line, 3 short paragraphs, then the question. No hashtags.
</format>

Ask me up to 3 questions first if anything important is missing.

Claude asked whether there was a price and a booking link, then produced a post that needed one small edit. The vague version produced something generic that would have needed a rewrite. Notice the constraint about thresholds: it stops the model from stating a number that may be out of date.

Hands-on: build your prompt template file

Create a document called Prompt library.
Add the C-T-C-F skeleton above as your master template.
Take the three prompts you use most often and rewrite each with C-T-C-F, tags and one example.
Run old and new versions side by side and note the difference in edit time.
Save the winners. In Module 3 you will turn the best of them into Project instructions or a Skill.

Pitfalls

Contradictory constraints: "detailed and comprehensive in under 40 words" forces the model to fail one of them.
Everything in one mega-prompt: break big jobs into steps (outline, approve, draft, critique, revise).
Negative-only instructions: "don't be boring" is weaker than "use one concrete example per paragraph".
Pasting secrets: passwords, card numbers and client confidential data do not belong in prompts; see Module 7.

How to measure success

Track two numbers for a fortnight: how many prompts produce a usable first draft, and how many minutes you spend editing. Structured prompts should move both in the right direction within days.

Video lecture: Chat fundamentals: the prompt structure experts use

Lecture coming soon · 15 chapters · about 9 minutes. Read the full transcript below.

Prompts that work
Why structure matters
Brief it like a new colleague
Context is the multiplier
Tags keep materials apart
Examples beat adjectives
Let Claude interview you
Iterate like an editor
Simple example
Worked example
Four prompt pitfalls
Try this now
Watch me do it, part 1
Watch me do it, part 2
Your next step

Lecture transcript

Prompts that work

Most people type a request into Claude the way they would type into a search box. Five words, fingers crossed. Experts do something different, and it takes about thirty seconds. In this lecture you will learn the four part structure professionals use, how to keep pasted material from getting mixed up, and how to iterate like an editor instead of pulling a slot machine lever.

Why structure matters

Why does structure matter so much? Because Claude cannot see inside your head. When you type write a post about our new service, you know the audience, the tone, the length and what your competitors are saying. Claude knows none of it. Every missing detail becomes a guess, and guesses average out to bland. Here is an analogy. Ordering from a tailor, you would not say make me a suit. You would give measurements, fabric, occasion and a deadline. A prompt is a set of measurements. The four part structure simply makes sure you never forget the important ones.

Brief it like a new colleague

Imagine briefing a talented new hire on their first day. You would never just say, write a post. You would say who it is for, what it needs to achieve, what the rules are and what the finished thing should look like. That is exactly the structure. Context, task, constraints and format. Claude's newer models follow instructions very precisely, so be direct. If you want suggestions only, say suggest, do not rewrite. If you want changes made, say rewrite. Precision in, precision out.

Context is the multiplier

Context does the heaviest lifting, and the most overlooked part of context is why. Telling Claude that procurement managers skim on their phones changes sentence length, structure and tone, all at once, without you writing three extra rules. When you explain the reason behind a constraint, the model can generalise it sensibly to situations you did not anticipate.

Tags keep materials apart

Next, tags. When you paste a brief, brand rules and three past posts into one message, wrap each in a labelled tag, such as brief, brand rules, past posts, written inside angle brackets. Claude is trained to respect that structure, so it stops blending details between documents. Even better, you can refer back precisely. Check every caption against the brand rules. It is a small habit with an outsized effect.

Examples beat adjectives

The strongest steering tool you have is an example. One or two short samples of what good looks like will beat a paragraph of adjectives like punchy or on brand. Just be explicit about what to copy, such as the structure, length and tone, and what not to copy, such as the specific product names or facts in the example.

Let Claude interview you

Here is the cheapest quality upgrade there is. End important prompts with this line. Before you start, ask me up to three questions if anything important is missing. Now, instead of guessing your price, your deadline or your call to action, Claude asks. A Manchester bookkeeping firm tried this on a LinkedIn launch post, and the model asked for the price and booking link, then produced a draft that needed one small edit.

Iterate like an editor

When the draft is close, give editor style feedback. Keep the structure, cut to one hundred and fifty words, remove paragraph two. When the draft is wrong because your prompt was missing something, do not stack five corrections. Edit your original message and resend it, so the context stays clean. Use retry only when the prompt was good and you simply want a different option.

Simple example

Here is a simple example you can try in one minute. Before, summarise this, with a supplier email attached. You get a paragraph that restates everything. After. Context, I manage purchasing for a small café chain and I read this on my phone. Task, summarise the attached email. Constraints, five bullets maximum, keep all prices and dates exactly as written. Format, bullets, then one line starting with risk, for anything that could cost us money. Now you get five tight bullets and a risk line pointing out that the delivery charge doubles below a minimum order. Same email, same model, ten extra seconds of typing.

Worked example

Let's see the difference on a real job. A six person bookkeeping firm in Manchester launches a monthly VAT readiness check. The vague prompt, write a LinkedIn post about our new service, produced something so generic it needed a rewrite. The structured version gave the audience, UK Shopify and Amazon sellers who worry about registration thresholds, set a length of one hundred and twenty to one hundred and sixty words, banned hype words, and asked for one question to invite comments. Notice one clever constraint. Do not state tax thresholds or rates, say check the current HMRC threshold instead. That single line stops the model from confidently quoting a number that may have changed.

Four prompt pitfalls

Watch for four pitfalls. Contradictory constraints, like detailed and comprehensive in under forty words, force the model to fail one of them. Mega prompts that try to do everything at once work better as steps, outline, approve, draft, critique, revise. Negative only instructions like do not be boring are weak, so say what you want instead, such as one concrete example per paragraph. And never paste passwords, card numbers or confidential client data into a prompt. To measure progress, track two numbers for a fortnight, how many first drafts are usable, and how many minutes you spend editing.

Try this now

Try this now. Scroll back through your Claude history and find the last prompt you wrote in under ten words. Rewrite it with the four parts. Add who it is for and why it matters. State the task in one sentence. Add two or three constraints, length, tone, anything that must not change. Specify the format. Then add the final line, ask me up to three questions first if anything important is missing. Run the new version in a fresh chat and compare it with the old answer. Most people find the new draft needs half the editing. Save the rewritten prompt as the first entry in your prompt library.

Watch me do it, part 1

Let me rebuild the Manchester bookkeeping prompt live. I start with what most people type. Write a LinkedIn post about our new service. I delete it and build it properly. Context. I run a six person bookkeeping firm in Manchester serving e commerce sellers, and we have launched a monthly VAT readiness check for sellers turning over ninety thousand to five hundred thousand pounds. Task. Write a LinkedIn post announcing it. Constraints. One hundred and twenty to one hundred and sixty words, British spelling, no hype words, do not state tax thresholds, say check the current HMRC threshold instead, and end with one question. Format. A hook line, three short paragraphs, then the question, no hashtags. Finally I add, ask me up to three questions first if anything important is missing.

Watch me do it, part 2

Claude replies with three questions. What is the price, is there a booking link, and should the tone be formal or friendly. I answer in one line each, forty nine pounds a month, the link, and friendly but professional. The draft arrives. I check the constraints one by one. Word count, fine. No hype words, fine. And instead of a threshold figure it says check the current HMRC threshold, exactly as asked. One tweak. I ask, keep everything else, but turn the hook into a question. Done. Finally, I copy the whole structured prompt into my prompt library under LinkedIn launch posts, so next time I only change the service details.

Your next step

Let's recap. Context, task, constraints and format. Tags to separate materials. Examples to steer. An invitation to ask questions. And editor style iteration. Your next step is in the lesson text. Start a prompt library, rewrite your three most used prompts with this structure, and compare how long you spend editing the results. Keep the winners, because in the Projects module you will turn them into standing instructions.

Key takeaways

Use Context, Task, Constraints and Format (C-T-C-F) for important prompts.
Wrap pasted materials in labelled tags so Claude can tell them apart.
Iterate with specific editorial feedback or edit the original prompt instead of regenerating blindly.
Invite Claude to ask clarifying questions before it starts.

Try it

Take a prompt you used recently and rewrite it using C-T-C-F plus tags. Run both versions and compare the outputs side by side.