Prompting Claude: structure and step-by-step thinking

Start from the basics you already have

Everything from the Foundations lesson on prompting still applies here: say who the answer is for (role), state the task clearly, give the context, and name the format you want. Treat the first reply as a draft to improve, not a verdict. If those four habits are new to you, that lesson is the place to start; they matter more than any Claude-specific trick.

This lesson adds two habits that Claude responds to especially well, and a note on the thing Claude is best at: taking in a lot of context at once.

Habit one: give Claude structure

The single most common muddle when working with a document is that Claude can't tell where your instruction ends and the material begins. You paste a policy, add "summarise the key points" at the bottom, and it all runs together as one blob of text. Claude usually copes, but you make its job far easier, and its answer far more reliable, when you separate the two clearly.

The fix is simple: label the parts. Put your instruction first, then mark the document with an obvious heading like "DOCUMENT:" or a line of dashes, so it's unmistakable which bit to act on and which bit to work from.

Tell me about this and what people need to do. Fernway operates a hybrid model with a minimum of two days a week in the office, core hours are 10:00 to 16:00, fully remote is by exception...

You are helping a new Fernway employee understand a policy. Read the document below and give me: (1) a three-sentence plain-English summary, then (2) a short bulleted list titled "What I need to do" of any actions the policy expects of an individual employee.

DOCUMENT:
[paste the hybrid-working policy here]

The second prompt does three structural things: it sets a role ("helping a new employee"), it names an exact output shape (a summary, then a titled list), and it fences off the document with a "DOCUMENT:" label. You'll get a cleaner, more consistent answer, and when you reuse the pattern next week with a different policy, you only swap the pasted text.

Habit two: ask Claude to think step by step

For anything that involves reasoning (comparing options, checking whether something adds up, working through a multi-part question) Claude does noticeably better when you ask it to work through it step by step before giving its answer. Left to rush, it can leap to a confident conclusion and reason backwards to justify it. Asked to lay out its thinking first, it catches its own mistakes.

Take a real Fernway wrinkle. The meeting notes say the office lease auto-renews unless notice is given by 31 August, and separately that a summer social is provisionally booked for 22 August. Suppose you want to know whether the team can safely leave the lease decision until after the social.

Below are our meeting notes. Question: can we safely leave the office-lease decision until after the summer social on 22 August? Work through the relevant dates and dependencies step by step first, then give me a clear yes/no with the reason. If the notes don't contain enough to decide, say so.

DOCUMENT:
[paste the meeting notes]

Claude will lay out the notice deadline (31 August), the social (22 August), and the dependency (a cost comparison is needed before deciding), then conclude. Because the reasoning is visible, you can check each step rather than trusting a bare "yes". That last line, "if the notes don't contain enough to decide, say so", gives Claude explicit permission not to guess, which is one of the most useful things you can add to any prompt.

Asking for step-by-step reasoning is especially worthwhile with the most capable model, which is built for exactly this kind of careful, multi-step thinking. For a quick summary you don't need it; for a judgement you'll act on, it's well worth the extra few seconds.

Habit three: hand Claude long context, and aim it

Claude's headline strength is that it can hold a lot of text in view at once, the context window, so you can give it whole documents rather than snippets. But "give it everything" works best when you also tell it where to look and what to ignore.

Three practical moves:

• Give the whole document, not a guessed extract. If you paste only the paragraph you think matters, Claude can't see the clause three pages later that changes everything. When in doubt, give it all and point it at the part you care about.
• Aim the question. "In the section on data protection, what does it say about AI tools?" is better than "what does this say about AI?", since it tells Claude where to concentrate.
• Ask it to ground its answer in the text. "Answer using only the document; if it isn't covered, say 'not covered'." This keeps Claude from topping up gaps with general knowledge that may not apply to your situation.

Using only the policy below, answer this: if I want to work fully remotely, who do I need to get agreement from, and how long should I expect to wait for a decision? Quote the exact lines. If the policy doesn't say, reply "not covered".

DOCUMENT:
[paste the full policy]

Putting it together

The three habits stack. A strong document prompt gives Claude a role and a clear output shape, fences off the material with a label, asks it to reason step by step where judgement is involved, and restricts it to what the document actually says. You don't need all of that for "summarise this", but the moment the task carries any weight, reaching for these turns a decent answer into a dependable one.

Try it now

Common mistakes

• Letting instruction and document blur together. Label the material with "DOCUMENT:" or dashes so Claude knows what to act on and what to work from.
• Asking for a judgement with no room to reason. For anything that needs working out, add "step by step first, then answer"; a bare conclusion hides its own mistakes.
• Pasting a guessed extract instead of the whole thing. Claude can't see the clause you left out. Give it the full document and aim the question.
• Forgetting to allow "not covered". Without permission to decline, Claude may fill a gap with plausible general knowledge that doesn't fit your document.
• Over-trusting a step-by-step answer because the reasoning looks thorough. Neat, confident-looking working can still rest on a misread line. The reason you ask for the steps is so you can check them: actually read them, and confirm any quote against the original, rather than being reassured by their mere presence.

Keeping current

The structure-and-reasoning habits are durable; they'll still work whatever the model is called. What shifts is how much Claude can take in at once and how it exposes its thinking (features like an "effort" control change over time). For current guidance on getting the best from a model, see Anthropic's choosing a model guide and the Claude release notes. Accurate as of 13 July 2026.