The one idea
There are no secret prompts.
The phrases that circulate as magic words — act as an expert, this is very important to my career — are at best mild nudges. What consistently works is unglamorous: say what you want clearly, give the material, set the constraints, show an example, and break big jobs into steps.
The part nobody teaches is what to do when the output is wrong. Most people reroll — same prompt, hoping for a better draw. That teaches you nothing. Diagnosing which of five things went wrong, and changing that one thing, is the skill.
Be specific, give it the material, say what good looks like, and show an example if you can.
Escalate deliberately: instruction, then context, then constraints, then examples and format, then a multi-step workflow with checkpoints. Only go up a level when the level below fails.
Each level further constrains generation. Examples are the strongest signal available for tone and format, because they demonstrate rather than describe. Structural separators between instructions and material reduce the chance your content gets read as instruction.
The five levels
Do not start at level five. Start at the lowest level that could work and climb only when it fails — otherwise you spend ten minutes engineering a prompt for a task that needed one sentence.
Level 1 — Direct instruction
Say exactly what you want. Most failures here are vagueness, not insufficient technique.
Weak: "Improve this email." Better: "Shorten this email to under 100 words and remove the apologetic tone."
"Improve" means nothing, so the tool guesses your definition — and it guesses "longer and more formal" surprisingly often.
Level 2 — Add context
Who it is for, what the situation is, what has already happened.
"This is going to a client who has already complained twice about delays. They are not technical. We are two weeks late and it is our fault."
That paragraph changes the output more than any clever phrasing will.
Level 3 — Add constraints
Boundaries, especially negative ones.
"Under 120 words. No bullet points. Do not offer a discount. End with a specific commitment and a date."
Negative constraints are underused and disproportionately effective: they rule out the average answer, which is exactly what you keep receiving.
Level 4 — Examples and structured output
An example does what a paragraph of description cannot: it demonstrates.
"Here are two emails I have sent that landed well. Match this register."
For structure, ask for the shape you will actually use. "Be concise" is a wish. "One table, four columns: Risk, Likelihood, Impact, Owner" is a specification you can paste straight into a document.
Also: separate your instructions from your material with clear markers. Otherwise a sentence inside your document that reads like an instruction can be followed as one.
Level 5 — Multi-step workflow with checkpoints
For anything substantial, do not ask for the finished thing. Break it up and inspect between steps.
"Step 1: read the transcript and list every commitment anyone made, with who made it. Stop and show me that list. Do not proceed until I confirm."
Then step 2 works from a list you have already corrected. One big request instead means an early error is baked silently into the final output, where you cannot see it.
The refinement table
This is the part to keep. When output disappoints, the cause is nearly always one of five things, and each has a different fix.
| What is wrong | What it means | What to change |
|---|---|---|
| Generic, could apply to anyone | It filled gaps with averages | Add context and specifics — who, where, what already happened |
| Right content, wrong tone or length | Constraints were absent or too soft | Add hard limits and an example of the register you want |
| Wrong shape — prose when you wanted a table | You described the output instead of specifying it | Name the exact structure, columns, sections |
| Confidently wrong facts | You asked it to produce what it cannot know | Change the job: supply the source material, or use it to find where to look |
| Started well then drifted | Task too big for one pass | Split into steps with a checkpoint after each |
You ask for meeting minutes from a transcript and get a readable summary that misses two decisions and invents a deadline nobody said.
Rerolling gets you a different summary with different gaps. The diagnosis is row five — too big for one pass — plus a verification problem. The fix: step one, extract every decision and every action item verbatim with the speaker's name, and nothing else. You read that list against the transcript. Step two, write the minutes from the corrected list.
The output is checkable, because you inspected the extraction before it became prose.
You want a first draft of a proposal section. It comes back plausible and bland — row one. You add the client's actual constraint ("they have an existing system they will not replace"), the real budget ceiling, and one paragraph from a proposal that won. The next draft is specific enough to edit rather than rewrite. The change was information, not phrasing.
Refining instead of rerolling
1 of 6Name what is wrong in one phrase. Too generic. Wrong tone. Wrong format. Factually wrong. Lost the thread. Force yourself to pick one — vague dissatisfaction cannot be fixed.
Try this
Someone asks for feedback on a two-page CV. They send this prompt:
"Act as an expert recruiter with 20 years of experience and give me detailed professional feedback on my CV to make it stand out."
The output is a long list of generic advice: use action verbs, quantify achievements, tailor to the role.
What would you change, and why?
Your challenge
Level 3 · IndependentTake a task where your first AI attempt disappointed you. Do not reroll it.
Diagnose the fault using the table, apply exactly one fix, and record the result. Then apply a second fix and record that. Keep going until it is usable or until you have made four changes.
You have succeeded when you can say which single change made the largest difference, and explain why in terms of what information the tool was missing.
If your answer is "I gave it more context and it got specific", you have learned the most transferable thing in this lesson — and you will reach for it first next time instead of last.
What people usually get wrong
- Collecting prompts instead of building judgement. A prompt written for someone else's task carries their context, which is the part that mattered.
- Rerolling the same prompt. Randomness means you eventually get a better draw. You have learned nothing and cannot repeat it.
- Piling on adjectives. "Detailed, comprehensive, professional, insightful" produces length, not quality.
- Describing format instead of specifying it. "Well structured" is a hope; named sections are an instruction.
- Not separating instructions from pasted material. Text inside a document can be read as a command to follow.
- Going straight to level five. A multi-step workflow for a one-line task wastes more time than the task.
- Believing role-play does heavy lifting. "Act as a lawyer" changes register and vocabulary. It does not add legal knowledge or reliability.
How someone experienced does it
Experienced users treat the prompt as a specification they iterate on, and they keep the specifications that survive. Over a few months they end up with five or six framing patterns for the work they actually do — not downloaded, grown.
Two habits worth stealing:
They ask the tool to critique the prompt before running it. "Here is what I am about to ask you to do. What is ambiguous about it, and what would you likely get wrong?" The answer is usually a short list of genuine ambiguities you can fix in twenty seconds, before spending a round on output you will discard.
They ask for the reasoning separately from the answer. "Give me your recommendation, then, separately, the three assumptions it rests on and what would change it." The assumptions are where the errors are visible — a recommendation reads as equally sound whether its foundations are solid or invented, but a stated assumption can be checked against what you know.
And they notice the ceiling: prompting improves how well a task is specified. It does not give the tool information it does not have. When the failure is missing information, no amount of rephrasing helps — you have to supply the material or change the job.
When not to use this
Elaborate prompting is the wrong move when the real problem is that the tool cannot do the task. If it needs a fact it does not have, access to a system it cannot reach, or judgement about your organisation's politics, a better prompt will produce a better-written wrong answer.
It is also the wrong move when you could do the task faster yourself. Fifteen minutes engineering a prompt for a four-minute job is a common and invisible way to lose an afternoon.
Why examples work better than descriptions
Describing a tone requires you to name it correctly and requires the tool to share your definition of the name. "Professional but warm" means something different to everyone, and the average of all those meanings is what you get.
An example carries the register, the sentence length, the vocabulary, the degree of directness, and a dozen things you could not name — all at once, without either of you having to agree on a word for them. This is why vendor prompting guidance consistently ranks examples among the most effective techniques available.
The practical consequence: build a small folder of your own good outputs. Two emails that landed well, one summary your manager liked, one document in your house style. Pasting one in as a model is faster than any description you could write, and it is genuinely yours.
Prove it
Pick one recurring task. Run it four times over a week, refining once per run, and log four lines each time: the prompt, what was wrong, the single change you made, and whether it helped.
The log is the deliverable. At the end you will have a prompt that works for your task and, more usefully, a record of which kinds of change fix which kinds of failure.
Keep learning this
Paste this into any AI assistant. It turns the assistant into a tutor that tests you instead of just answering you.
Act as an experienced practitioner who is good at teaching. I have just learned refining a prompt when the AI output is wrong. Assume I am intelligent but relatively new to this — treat me as intermediate level. Work through this in order, and wait for my reply at each step: 1. Ask me 5 questions that test whether I actually understood refining a prompt when the AI output is wrong. Do not reveal the answers yet. 2. After I answer, tell me which parts I got right, which I got wrong, and which I only half-understand. Explain only what I misunderstood — do not re-teach what I already know. 3. Give me one practical challenge based on something I could genuinely encounter at work or in daily life. Do not solve it for me. 4. Evaluate my solution the way an experienced person would judge it, including what a professional would have done differently. 5. Tell me what to learn next, and why that comes next. 6. Give me trustworthy sources for deeper study — prefer official documentation, primary research or standards bodies over blogs and videos. Rules for you: no buzzwords. No motivational filler. Say "I'm not certain" when you are not certain, and tell me which parts of your answer I should verify myself. Clearly separate facts from your recommendations and your opinions.
Become independent at this
Use this when you want a path from where you are to actually good, with checkpoints you can test yourself against.
I want to become independently capable at getting usable output from an AI tool — not permanently dependent on AI, tutorials or step-by-step guides. Design a progression for me with five stages: Beginner, Guided practice, Independent practice, Real-world application, Professional level. For each stage tell me: - what I must know - what I must be able to do without help - the mistakes people make at this stage - one practical challenge - one real project that would prove I reached this stage - one way I can test myself honestly Then tell me the signals that I am ready to move to the next stage, and the signals that I have skipped ahead too early. Keep the theory to the minimum I actually need. Focus on ability I can transfer to situations you and I have not discussed.