Brief Claude like a new employee
Anthropic's prompting guide offers one comparison that is worth remembering: treat Claude as a brilliant new employee who does not know your norms or workflows, and explain precisely what you want1. Its golden rule is to show the prompt to a colleague with little context, and if they would be confused, Claude will be too1.
That framing fixes most weak prompts. A request like write a reply to this complaint gives Claude nothing about the customer, the history, your policy or your tone, so it fills the gaps with generic text. The same request with three sentences of context usually produces a draft you can send after light editing.
Explaining why also helps. The guide gives the example of telling Claude not to use ellipses because the text will be read aloud by a text-to-speech engine, and notes that Claude generalises from the reason1. In business terms: say the email is going to a farmer reading on a phone in the shed, and Claude will keep it short without being told to.
The six parts of a strong prompt
Not every prompt needs all six, but the ones that matter for client work usually do. Write them in plain sentences. You do not need special syntax in the Claude app.
- Role: who Claude is acting as, in one line. For example, an experienced office manager at a Christchurch building company.
- Context: the reader, the purpose and the situation, including anything a newcomer would get wrong.
- Source material: the actual document, thread or spreadsheet, uploaded or pasted, rather than your summary of it.
- Task: what to produce, as numbered steps if order matters. Anthropic recommends sequential steps when completeness or order is important1.
- Format: length, structure, spelling and tone. Say what you want, not only what you do not want; the guide suggests asking for flowing prose rather than saying do not use markdown1.
- Examples and limits: one to five examples of good output, and permission to say it does not know. Anthropic suggests 3 to 5 varied examples for the best results1, and says explicitly allowing Claude to admit uncertainty can sharply reduce made-up answers2.
Long documents: put the document first and ask for quotes
Claude is strong on long material, which is where many New Zealand firms get their first real win: leases, tender documents, council consent conditions, insurance policy wordings. On paid plans the context window for the current main models runs to 1M tokens in chat, and older models have at least 200K, which Anthropic describes as about 500 pages of text or more3.
Two techniques from Anthropic's guidance make long-document work more reliable. Put the long material near the top and your question at the end; Anthropic reports that queries placed at the end improved response quality by up to 30 percent in its tests on complex, multi-document inputs1. And ask Claude to pull out the relevant quotes first, then answer using only those quotes1,2.
Know the file limits. In a chat, Claude analyses both text and visuals in PDFs of up to 100 pages, and reads text only for PDFs between 101 and 1,000 pages4. When you refer to a page, use the page number your PDF viewer shows, not the number printed on the page4.
Worked example: a commercial lease for a Wellington retailer
A retail owner is renewing a shop lease and wants to understand the obligations before meeting the landlord's agent. Upload the lease, then use a prompt like this.
Prompt: You are helping the owner of a small Wellington clothing shop prepare for a lease renewal meeting. The lease is attached. First, list word-for-word quotes for every clause about rent reviews, outgoings, repairs and maintenance, make-good at the end of the lease, and the right to renew, with the clause number for each. Then, using only those quotes, write a one-page summary of what the tenant must pay or do, and a list of five questions to ask the agent. If the lease does not cover something, say not covered rather than guessing. Use New Zealand spelling. This is preparation for a meeting, not legal advice, so flag anything that a lawyer should look at.
Why it works: the role and situation are specific, the quote-first step keeps the summary tied to the document, and the instruction to say not covered stops Claude filling gaps. The owner still checks every quote against the lease before the meeting, and the lawyer still reviews the final terms.
Worked example: a customer reply for a Queenstown tourism operator
A jet boat operator gets an angry email after a trip was cancelled for river conditions. The cancellation policy is in the business's Claude Project, so the prompt can refer to it.
Prompt: Draft a reply to the email below from a customer whose trip was cancelled this morning because of high river levels. Our cancellation policy is in the project knowledge; apply it exactly and quote the relevant line. The customer is upset and on holiday with family, so the tone should be warm and direct, not defensive. Offer the two options the policy allows. Keep it to two short paragraphs of plain text with no bullet points, signed by the duty manager's first name. Here is a reply we sent last month that had the right tone: then paste the example.
Why it works: the example sets the tone better than any adjective, the policy is quoted rather than paraphrased, and the length and format are fixed. Before sending, the duty manager checks the options against the actual policy and the booking system.
Worked example: job costing for a Canterbury builder
A small residential builder exports job costs from their accounting system as a CSV and wants to see which job types overrun. Claude can run code to analyse data and create Excel files with working formulas, provided code execution and file creation is turned on5.
Prompt: Attached is a CSV of completed jobs from the last two financial years, with quoted price, actual cost, job type and suburb. First, describe the columns and flag any rows that look like data errors, and do not change them. Then calculate the overrun for each job as actual cost minus quoted price, and summarise the average overrun by job type. Create an Excel file with the cleaned data on one sheet, the summary on another, and formulas rather than pasted values, so I can check the working. Finally, give me three plain-language observations, and say where the data is too thin to draw a conclusion.
Why it works: it separates checking the data from analysing it, asks for formulas so the owner can audit the numbers, and asks Claude to say where the evidence is weak. Open the spreadsheet and test a few rows by hand before relying on it.
Two more patterns worth keeping
Tender and RFP responses. A Hamilton engineering consultancy answering a regional council request for proposal can upload the RFP and its own capability statement, then ask Claude to build a compliance matrix listing every mandatory requirement with the RFP section number, and mark which ones the capability statement already answers. That turns a long read into a checklist the bid lead can work through.
Research with sources. For questions that need current information, such as changes to a supplier's published terms, turn on Research from the plus menu. It runs several web searches that build on each other and returns answers with citations, and it needs web search turned on6. Anthropic still advises reviewing the cited sources, because the original pages may contain context that the summary missed7.
Save good prompts so the team can reuse them
A prompt that works once should not live in one person's chat history. There are three places to keep it, depending on scope.
Account-wide habits, such as always using New Zealand spelling, belong in the personal Instructions for Claude setting, which applies to all of a person's conversations8. Context and rules for one kind of work belong in project instructions, which apply only to chats inside that Project8. Repeatable procedures, such as how to format a site report, are a good fit for a skill: a packaged set of instructions Claude loads only when the task needs it, which Team owners can provision for everyone9.
Finally, match the effort to the task. The model menu next to the send button sets an effort level; lower effort suits routine tasks and stretches usage further, while higher effort gives more thorough answers but takes longer and uses more of your limit10. A quick email needs little. A lease review deserves more.
- Check every fact, figure and quote against the source before it leaves the business.
- Start a new chat for a new task, so old context does not leak into the answer.
- When an answer is weak, add the missing context and ask again, rather than arguing with the draft.