Kape Tools

Prompt Engineering

Writing the instruction so the answer comes back right

Prompt engineering is just writing your request well. It is the cheapest skill here and the one that changes your results the most. No coding involved.

What it really is

Prompt engineering sounds technical. It is not. It means writing your request clearly enough that the AI can only give you one kind of answer: the one you wanted.

Most people blame the AI when the answer is generic. Nearly always, the request was generic first.

Why the AI cannot ask you

A human freelancer would message you back and ask. The AI usually will not. It answers with what it has. So a missing detail does not come back as a question. It comes back as a guess that looks confident.

A weak prompt on the left producing a vague answer, and a strong prompt on the right built from four parts, producing a usable answer
Same job, same AI. The only thing that changed is the instruction.Open full size

The four parts

  1. Context. Who you are and what this is for. "I run a small VA agency in Cebu."
  2. The task. One clear thing to make. "Write a follow-up email to a client who has gone quiet."
  3. Constraints. Length, tone, format, and what to never do. "Under 120 words. Warm but not desperate. Do not offer a discount."
  4. What good looks like. An example, or a standard to match. "Match the tone of the message I pasted below."

You will not always need all four. You will notice fast when you skipped the one that mattered.

See the difference

Weak:
  Write a proposal for a client.

Strong:
  I do social media management for small restaurants.
  Write a proposal for a client who asked for help after
  their posting went quiet for 3 months.

  One page. Three deliverables, no more.
  Use the prices in the rate card I pasted below.
  If a price is not in there, write [CONFIRM] instead of guessing.
  End with one next step and a date, never "let me know".

The strong version takes about thirty seconds longer to type. It saves three rounds of "no, not like that".

Habits that work straight away

The one line that saves you most

Ask for thinking, not just an answer

When there is a decision involved, ask for options and the trade-off of each. "Give me three options and what each one costs me" gives you something you can judge. "What is the best option" gives you one confident answer with no way to check it.

Mistakes that quietly waste your time

MistakeWhat happensDo this instead
Being polite instead of clearPadding, no extra qualityBe direct. It does not get offended.
Five questions in one messageFive shallow answersOne task per message
One chat running all weekAnswers get vaguerNew chat per task
Trusting a number because the writing was goodA wrong figure in a client documentCheck anything you would be embarrassed to get wrong

When a better prompt is not the answer

Prompt engineering has a ceiling. Better wording cannot give the AI a fact it was never shown, cannot empty a conversation that is already full, and cannot hand it an ability it does not have.

Four things that look like bad answers, matched to what actually fixes each one: only the first is a wording problem, the other three need a file, a fresh chat, or a tool
Four failures that look identical from the outside. Only the first one is about wording.Open full size

Most of what gets called "bad AI" is one of the last three. Rewriting the question a fourth time is the most common way to lose an afternoon to a problem that was never about the question.

Where prompts should live

If you type the same instructions every day, stop typing them. Put them somewhere they load by themselves: a saved style, a project instruction, a skill file, or a CLAUDE.md. A prompt you rewrite daily is a prompt you have not finished writing.

A one-off request retyped every time, turned into a saved template where only the highlighted details change, and four places it can live so it loads by itself
Same words. The difference is that the second one loads itself.Open full size
  1. Take the one that worked. Not the one you meant to write, the one whose answer you actually used.
  2. Swap the specifics for slots. Client name, situation, price. Everything else stays exactly as it was.
  3. Save it where it loads by itself. A saved style, project instructions, a skill file, or a CLAUDE.md.
  4. Fix it in that one place when it lets you down, so tomorrow's version is better than today's.

More words

See all 11 explanations