Prompts that work the first time

In module 2 you saw that AI only knows what is in its context window, and in module 3 you picked a main tool. This module is the Description skill from the 4D framework in depth: how to ask for things so the first draft comes out right. "The first time" does not mean perfect; it means what comes back is 80 percent there and the fix takes two messages instead of twenty.

We will cover five things: the parts of a good instruction, the techniques Anthropic documents for its models (which work almost identically in ChatGPT or Gemini), how to correct a draft, how to turn the prompts that worked into a company asset, and the mistakes that come up again and again.

What are the parts of a prompt that works?

Almost every instruction that comes out well has the same seven pieces. You will not always need all of them, but every one you leave out is a decision the model makes for you, and it makes it with what it saw on the internet, not with what you know.

1. 1Role. Who it should be when it answers. "You are the sales manager of a food distribution company" changes the vocabulary and what it treats as important.
2. 2Context. The real situation: what happened, with whom, what is at stake. This is where the most is gained and what people skip most often. Module 2 was entirely about it.
3. 3Task. A clear verb and a concrete object: draft, summarize, compare, classify, propose three options. One main task per instruction.
4. 4Format. How you want the result: length, structure, language, tone, table or prose, whether it is going into an email or a slide.
5. 5Constraints. What it must not do and, better still, why. "Do not mention discounts, because management has not approved them yet."
6. 6Examples. A text of yours that you like, or a sample of the format you expect. Showing beats describing.
7. 7Success criteria. How you will know it came out right. "The supplier should understand that the relationship continues, but that the next incomplete delivery gets deducted from the invoice."

Compare these two versions of the same task, at a food company that just received an incomplete delivery from its packaging supplier.

• Weak: "Write an email to the supplier about the incomplete delivery."
• Strong: "You are the head of purchasing at a food company in Bogotá. Our packaging supplier, who we have worked with for three years without problems, delivered Monday’s order with half the boxes missing and no warning; the plant had to stop for half a shift. Write the email to their account manager. Goal: they replace what is missing this week, and it is clear that a second failure gets deducted from the invoice, without damaging the relationship. Professional, firm tone, no legal threats, 150 words maximum, in Spanish. Do not invent figures or dates I did not give you; if you need one, leave it in brackets. Here is an email I sent another supplier so you can copy the tone: [PASTE ONE OF YOUR EMAILS]."

The second one takes two minutes longer to write. In return, the draft arrives in the right tone, with no promises you cannot keep, and with the gaps marked instead of filled in. Notice that the context is not "we are a food company"; it is the specific story that explains why the email has to be firm but not hostile.

What techniques does Anthropic recommend for better prompts?

Anthropic maintains a public best-practices guide for its models. What follows is, as of October 2026, the part of that guide most useful to someone who does not program. It applies to Claude, and in practice to the other AI assistants too.

• Be clear and direct. Current models respond well to explicit instructions; if you want it to go beyond the basics, ask for that instead of hoping it infers it. The guide uses the same mental model as this course: a brilliant but new employee who lacks context on your norms.
• Explain the why behind each rule. According to Anthropic, giving the motivation behind an instruction helps the model understand your goal and generalize. Its example: "Never use ellipses" works worse than "Your response will be read aloud by a text-to-speech engine, so never use ellipses, since the engine will not know how to pronounce them." From that, the model works out everything else a voice reader cannot handle.
• Give examples. Anthropic describes them as one of the most reliable ways to steer format, tone and structure. Its recommendation: three to five relevant examples, varied among themselves and clearly separated from the instructions.
• Separate the document from the instructions. When you paste a contract, a set of regulations or an email thread, mark it with a clear section: a heading like "DOCUMENT:" and another like "INSTRUCTIONS:", or tags such as <document> and </document>, which is what Anthropic recommends so the model does not confuse what it should read with what it should do. For long texts, the guide suggests putting the document at the top and the question at the end.
• Ask for the format you want, in the positive. The guide puts it this way: tell it what to do instead of what not to do. "Write in flowing paragraphs" works better than "Do not use lists". And your prompt’s style rubs off on the answer: write in bullet points and you get bullet points back.

• Let it think on hard problems. As of October 2026, current Claude models reason before answering and calibrate how much to think based on the complexity of the question; Anthropic notes that this behavior can also be steered from the prompt. For a decision with several variables (whether to renew a contract, choosing between three suppliers) ask it to analyze the options and risks first and only then recommend. For "fix the spelling in this paragraph", there is no need.
• Give it permission to say "I don’t know". It is the first technique in Anthropic’s guide to reducing hallucinations (module 1): explicitly give the model permission to admit uncertainty, which according to Anthropic can drastically reduce false information. One line is enough: "If a fact is not in what I gave you, say so instead of filling it in." For long documents, the same guide suggests asking it to extract word-for-word quotes first and then work only from those.
• Ask it to ask you first. End with "Before you start, ask me whatever questions you need, five at most". The model spots what is missing (the date, the amount, who it is for) and asks instead of assuming. It is the cheapest way to discover which context you forgot.

I want to ask you for something but I cannot explain it well yet. Here is my idea, as it comes out:

[DESCRIBE WHAT YOU WANT, UNSORTED]

Before doing the task, follow these three steps:

1. Ask me the questions you need to do it well (6 at most, numbered). Ask above all about the context, who the result is for, and how I will know it came out right.

2. When I answer, write the complete instruction you would use yourself for this task, with these parts labeled: ROLE, CONTEXT, TASK, FORMAT, CONSTRAINTS (with the reason for each), EXAMPLES (if I gave any) and SUCCESS CRITERIA. Leave in [BRACKETS] whatever I would have to fill in next time.

3. Run it and give me the result. If something is not in what I gave you, leave it in brackets instead of inventing it.

How do I fix a draft without starting over?

The first draft is almost never the final one, and that is not a failure: Anthropic’s blog says it plainly, the first prompt rarely works perfectly, so test and refine. The skill is in how you correct. Rewriting the prompt from scratch throws away what already came out right; saying "make it better" gives it nothing to work with. Feedback that works looks like the feedback you would give a person:

• Say what to keep. "The structure and the second paragraph stay as they are." If you do not say so, it may change what was already fine.
• Say what to change, and in which direction. Not "shorter", but "100 words maximum, cut the thank-you paragraph". Not "friendlier", but "open by acknowledging the three years of good business".
• Say why. "Remove the mention of a penalty, because the contract does not include one." With the reason, it fixes that sentence and the similar ones.
• Ask it to check itself. "Compare your draft with the context I gave you and flag any fact or promise that is not in it." It catches its own inventions surprisingly often.

There is a point where correcting stops helping and a fresh conversation is better. You saw it in module 2: everything said in the chat stays in the context window, including the drafts you discarded. Signs that it is time to restart: the model mixes versions, goes back to a mistake you already fixed, or the task has changed from "write the email" to "now make me the presentation". Copy the best prompt and the best draft, open a clean chat and paste only those.

How do I turn a good prompt into a company asset?

A prompt that worked once is a result. A prompt that is saved, named and has its blanks marked is an asset: the next person who gets that task starts at 80 percent. We call that a {{prompt library|biblioteca-de-prompts}}, and it is built like this:

1. 1Turn the prompt into a template. Replace today’s specifics with [BRACKETS]: [SUPPLIER NAME], [WHAT WAS MISSING], [DEADLINE]. What does not change between uses (role, tone, constraints, the why) stays fixed.
2. 2Give it a name people can search for. Verb plus object plus version: "Supplier complaint email v2", "Regulations summary for the board v1". Add one line on when to use it and the date it was last reviewed.
3. 3Store it where the team can see it. A shared document (Google Docs, Notion, whatever you already use) with one section per area is enough to start. Not in one person’s chat history.
4. 4Move the fixed parts into a project. As of October 2026, Claude’s projects exist on every plan, including the free one, and hold fixed instructions and documents the model reads in every conversation; sharing them with the team requires a Team or Enterprise plan. ChatGPT and Gemini have equivalents. The role, tone and constraints that repeat across your whole library go there, with your context document from module 2. Module 6 covers this in depth.
5. 5Review the library every month. Every correction you made twice becomes a line in the template. Every template nobody used in three months gets archived.

This closes a loop that runs through the whole course. In the AI Fluency framework, Description and Discernment are not separate steps but a cycle: you describe, you evaluate what came out, you discover what the description was missing and you adjust it. A prompt library is that loop turned into a document. In part 4 you will see the same idea under another name: a specification for an agent is a prompt template with acceptance criteria.

ROLE: You are [WHO IT SHOULD BE: job title, experience, what kind of company].

CONTEXT: [WHAT HAPPENED, WITH WHOM, WHAT IS AT STAKE. Two to five sentences with the specific situation, not the general description of your company.]

TASK: [ONE VERB AND ONE OBJECT: draft, summarize, compare, classify, propose N options.]

FOR WHOM: [WHO WILL READ OR USE THE RESULT and what that person already knows.]

FORMAT: [MAXIMUM LENGTH, STRUCTURE, LANGUAGE, TONE. Example: "200 words maximum, three paragraphs, English, professional and warm".]

CONSTRAINTS:
- [WHAT NOT TO DO] because [WHY].
- Do not invent facts, names or figures that are not in this message; if something is missing, leave it in [brackets].
- If anything I am asking for is contradictory or cannot be done, tell me before you start.

EXAMPLE OF WHAT I LIKE (optional):
[PASTE A TEXT OF YOURS OR A SAMPLE OF THE FORMAT]

SUCCESS CRITERIA: It came out right if [HOW YOU WILL KNOW IT WORKS: the reaction you expect, what must be clear, what cannot be missing].

Before you start, if you are missing something important, ask me (three questions at most). Otherwise, go ahead.

What are the most common prompting mistakes?

Almost every prompt that fails does so for one of these five reasons.

• The one-line instruction. "Summarize this contract." For whom, to what end, how long, what worries you about the contract? Every unanswered question is a decision the model makes for you.
• Vague adjectives. "Make it more professional", "make it sound better", "more engaging". They describe nothing measurable. Replace them with the concrete thing you have in mind: "no exclamation marks, the price in the second sentence, formal address".
• Contradictory rules. "Be brief but cover every detail. Be formal but warm. No lists, but easy to scan." The model tries to satisfy everything and satisfies nothing. Anthropic warns that longer, more complex prompts are not automatically better; every rule you add should solve a problem you have already seen.
• Asking for too many things at once. "Summarize the regulations, propose changes, draft the email to the co-owners and make me the presentation." Four tasks means four prompts, or one conversation in four steps where you review each before moving on.
• Not saying who it is for. A summary for the finance manager and one for the fleet drivers are different texts even when the document is the same. If you do not name the reader, the model writes for nobody in particular, and it shows.

I am going to paste a prompt I use at work. Do not run the prompt: evaluate it.

PROMPT:
[PASTE YOUR PROMPT HERE]

What I use it for: [THE TASK AND WHO USES THE RESULT]
What frustrates me about the results: [WHAT COMES OUT WRONG OR UNEVEN, IF YOU KNOW]

Do this, in order:
1. Check which parts it has and which are missing: role, context, task, audience, format, constraints with their reasons, examples, success criteria.
2. Point out vague adjectives, contradictory rules, several tasks mixed into one, and any instruction a colleague without context would not understand.
3. Tell me what information you would need from me for the prompt to work better, as concrete questions.
4. Write the improved version, with the parts labeled and [BRACKETS] around whatever changes from one use to the next.

Be direct. I would rather have a harsh critique than a prompt that keeps failing.

That closes the Description skill. One piece of part 1 remains: before you paste contracts, customer lists and employee data into your prompts, you need to know what happens to that information and what Colombian law says about it. That is module 5, "Privacy, risk and judgment".