Contents
The principle: give a brief, not just ask a question
Anthropic's documentation on prompt engineering emphasizes a few simple rules: be clear and direct, provide context, include examples, structure the information, and let Claude think before answering. There's no magic: it's exactly what you'd do with a new collaborator.
The method in 6 elements
- Context: who you are, who the result is for, why you need it.
- Task: a precise action verb (write, summarize, compare, classify, correct).
- Data: the document, figures or notes to work from.
- Constraints: length, tone, what to avoid, language.
- Format: a table, a list, an email ready to send, an outline…
- An example: a model of the expected result, worth more than ten lines of explanation.
Before / After: three concrete examples
Example 1, an email to a client
Before:
Write a follow-up email to a client.
After:
I run a real estate agency in Madrid. Draft a follow-up email for a buyer who visited an apartment five days ago and hasn't responded. Tone: friendly and professional, without pressure. Maximum 120 words. Propose two times to call them and end with an open question. Give me two versions: one formal and one more direct.
Example 2, summarizing a document
Before:
Summarize this contract.
After:
You are in-house counsel. Read the attached contract and create a table with: each party's obligations, dates and deadlines, amounts, termination clauses, and any unusual or risky points. Finish with the 3 questions to ask the supplier before signing. If information is missing, write «not specified» instead of guessing.
Example 3, content in a given style
Before:
Write a LinkedIn post about our new service.
After:
Here are three LinkedIn posts I published that performed well: [paste]. Analyze their style (length, hook, tone, structure). Then write a post in that same style to announce our new service of [describe]. Audience: SME executives. One strong idea only, no strings of hashtags.
5 techniques that make a difference
Structure with tags. In a long prompt, separate parts: `<context>…</context>`, `<document>…</document>`, `<instructions>…</instructions>`. Anthropic recommends this practice because it prevents Claude from mixing instructions with the data.
Ask Claude to ask questions. Add: 'Before starting, ask me any questions necessary to do an excellent job.' This is ideal for complex tasks.
Allow reasoning. For an analysis or decision, ask it to reason step-by-step before concluding.
Allow the "I don't know." Indicate that it should flag missing information instead of inventing it. This greatly reduces errors.
Iterate. Don't start over from scratch: say exactly what is off ('shorter', 'less promotional tone', 'add an example with figures').
Save time: from the prompt to Projects
If you often rewrite the same context, put it in a Project's instructions. Claude will apply it to all conversations in that Project and your prompts will become short. Also save your best prompts in a document shared with the team.
Template to copy
Context: [who I am, who it's for, why] Task: [verb + precise objective] Data: [document, figures, notes] Constraints: [length, tone, what to avoid, language] Format: [table, list, email…] Example: [model of the expected result] If you lack information, ask me before starting.
Frequently asked questions
Do prompts have to be written in English?
No. Claude understands Spanish and French perfectly. Write in the language you expect the output to be in.
Does a prompt have to be long?
It needs to be complete, not necessarily long. For a simple task, two well-chosen sentences are enough. For an important task, a prompt of 10 to 20 lines is normal.
Why doesn't Claude give the expected result?
Most often context, a format or an example is missing. Add them before concluding that the tool can't do it.
Sources
- Anthropic, *Prompt engineering overview* (docs.claude.com)
- Claude Help Center, *How can I create and manage projects?* (support.claude.com)
