Most prompt lists fail for the same reason: a one-line question isn’t a prompt; it’s a topic. “Challenge my assumptions” gives the model nothing to work with, so it returns something that could apply to anyone. Below is the anatomy of a prompt that works, then twenty-two complete prompts arranged by where they sit in the content pipeline — with an honest note at each stage about what the model is actually good at.
Why one-line prompts return generic output
Current models take instructions literally and do what you asked, not what you meant. Ask a general question, get a general answer — the model has no way to know your audience, your angle, or what you’d consider useful. It fills the gaps with an average of everything it has seen.
A working prompt supplies four things:
| Part | What it does | Missing it causes |
|---|---|---|
| Task | The action, stated with a direct verb | Model guesses what you want done |
| Context | Audience, purpose, what you already have | Output aimed at nobody in particular |
| Constraints | What to include, exclude, prioritise | Padded, unfocused, wrong length |
| Output shape | Format, structure, how you’ll use it | Prose when you needed a list, or the reverse |
Two additions that reliably improve results and cost one line each. Explain why a constraint exists rather than just stating it — models reason about the underlying goal and make better judgement calls on the cases you didn’t anticipate. And give explicit permission to admit uncertainty, which reduces confident invention.
Techniques worth dropping
- “Act as an expert.” A persona alone adds nothing current models don’t already do. A specific role in a system prompt (“senior technical editor reviewing for a developer audience”) still helps; “act as a writing expert” prepended to a chat message does not.
- “Think step by step.” Reasoning models do this internally. Adding it manually is redundant, and phrases like “think carefully about this” don’t move the dial — they just consume tokens.
- Long always/never lists. State what to do rather than stacking prohibitions. “Use natural paragraphs” beats “never use bullet points, never use headers, never use…”
- Piling on examples. Current models attend closely to examples and will reproduce their patterns, including ones you didn’t intend. Start with one; add a second only if output still misses.
Stage 1 — Interrogate the topic
Good for: mapping what exists before you commit. You supply: the audience. Without it every answer describes a generic reader.
I'm writing for [AUDIENCE] about [TOPIC].
List 15 questions this audience actually asks about it, then mark each:
- WELL COVERED — existing articles answer it properly
- POORLY COVERED — answers exist but are generic, dated, or contradictory
- UNANSWERED — people ask and nobody addresses it directly
For the poorly covered and unanswered ones, add one line on what's missing.
Where you're not confident how well something is covered, say so rather than guessing.[AUDIENCE] are learning [TOPIC].
List the 10 mistakes they make most often. For each: what they do, why it
seems reasonable at the time, and what it costs them.
Order by how expensive the mistake is, not how common.Summarise the main positions people hold on [TOPIC].
For each: who holds it, their strongest argument, and the weakest point
critics attack.
Where a disagreement is really about definitions rather than substance, say so.
Where the evidence genuinely settles it, say that too.Stage 2 — Find the angle
Good for: volume and variation. You supply: the judgement about which angle you can actually write. The model will happily hand you twenty angles you have no standing to argue.
Topic: [TOPIC]
Audience: [AUDIENCE]
Generate 20 angles. Vary the approach across:
- A contrarian take on received wisdom
- A specific failure and what it taught
- A comparison nobody makes
- A process nobody documents
- A definition people get wrong
Mark any angle that requires first-hand experience to write credibly.I want to write about [TOPIC], but every version I've found is a tips listicle.
Reframe it as: what goes wrong, why it happens, and what to do instead.
Give me 10 framings in that shape.
Skip anything that's a tips list wearing a different hat.My angle: [ANGLE]
Audience: [AUDIENCE]
Answer three questions:
1. What would a reader have to already believe for this to feel new?
2. Who would disagree, and what's their best counter?
3. What existing article makes the closest case, and where does mine differ?
If the angle is essentially common knowledge, say so plainly.Stage 3 — Structure
Good for: spotting gaps in a plan. You supply: the angle. An outline built without one produces the same shape as every other article on the topic.
Working title: [TITLE]
Audience: [AUDIENCE]
Angle: [ANGLE]
Build an outline ordered by what the reader needs next, not by the order I'd
naturally explain it.
For each section: the heading, the one thing it must establish, and roughly
how much space it deserves relative to the others.
Flag any section that exists only because articles like this usually have one.Here's my outline:
[PASTE OUTLINE]
What's missing that the reader needs? Consider:
- Steps I've assumed but never stated
- Terms used before they're defined
- Obvious objections left unaddressed
- The question they'll have immediately after finishing
List real gaps only. Don't pad it to reach a number.Stage 4 — Draft
Good for: less than you’d hope. This is the stage where delegation costs you the most — model prose is competent and anonymous, and editing it into your voice usually takes longer than writing it. Use the model to get unstuck, not to produce text.
I'm stuck on one section. Surrounding context:
[PASTE THE SECTIONS BEFORE AND AFTER]
This section needs to establish: [WHAT IT MUST DO]
Don't write it. Give me three different approaches, one line each, and note
what each costs in length and reader attention.This passage explains [CONCEPT]:
[PASTE PASSAGE]
My current example is [EXAMPLE] and it isn't landing.
Suggest five alternatives drawn from contexts [AUDIENCE] already understands.
For each, note what it captures well and where it distorts the concept.Stage 5 — Revise
Good for: more than any other stage. Diagnosis is where models genuinely outperform a tired author. You supply: the decisions — every prompt here reports rather than rewrites, deliberately.
Here's my draft:
[PASTE DRAFT]
Mark every passage that could appear in any article on this topic by any
author. Quote the passage, then say what specific claim, number, or first-hand
detail would replace it.
Don't rewrite anything. Just show me where the article is empty.[PASTE DRAFT]
Find where I make the same point more than once in different words.
For each cluster: quote the instances, and say which is strongest.[PASTE DRAFT]
Audience: [AUDIENCE]
What are the three strongest objections a knowledgeable reader would raise?
For each: the objection, where in the draft it arises, and whether I address
it, dodge it, or miss it entirely.[PASTE DRAFT]
This runs [CURRENT] words and needs to be [TARGET].
Identify what to cut, in priority order, with the reason for each.
Preserve: [WHAT MUST STAY].
Don't produce a shortened version. Give me the cut list so I decide.Stage 6 — Adapt to other channels
Good for: reformatting something that already works. You supply: nothing extra — but note that “turn this into ten posts” fails because most of an article doesn’t survive removal from its context. Extract first, then adapt.
[PASTE ARTICLE]
Identify the points that stand alone — each has to make sense cold to someone
who hasn't read the article.
For each: the claim in one sentence, and which format suits it
(short post, thread, carousel, video hook).
Skip anything that only works in context.Claim: [ONE CLAIM FROM THE ARTICLE]
Platform: [PLATFORM]
Audience: [AUDIENCE]
Write three versions. Vary the opening move: one leads with the
counterintuitive part, one with a concrete number or detail, one with the
failure it prevents.
No hashtags, no engagement questions, no "here's the thing."Article: [SUMMARY]
What someone types to find it: [SEARCH PHRASE]
Give 10 title options. Each must describe what's actually in the article —
no curiosity gaps, no "the X nobody tells you."
For each, note the search phrasing it targets.Article: [TITLE AND SUMMARY]
Audience: [AUDIENCE]
What five articles would surround this one to form a coherent cluster?
For each: the title, the specific question it answers that this one doesn't,
and which direction the link should run.
Skip anything that would substantially overlap with this article.Stage 7 — Pressure-test
Good for: finding what you can’t see because you wrote it. You supply: the willingness to act on the answer. These only pay off if you’re prepared to hear that the piece isn’t ready.
[PASTE ARTICLE OR THESIS]
Make the strongest case against this — the version a smart person who
disagrees would actually argue, not a strawman.
Then tell me which parts of that case I can't currently answer.[PASTE DRAFT]
List the assumptions I treat as settled without arguing for them.
For each: is it uncontroversial, contested, or wrong? For the contested ones,
say what I'd need to add to earn it.[PASTE ARTICLE]
Audience: [AUDIENCE]
If someone sent this to a colleague, what sentence would they write in the
message?
If you can't construct one that isn't generic, the article doesn't have a
point worth sharing yet. Say that.Business decisions
Good for: surfacing options you hadn’t considered. You supply: the observed evidence. A model asked to invent your audience’s problems will invent plausible ones that nobody has.
Audience: [AUDIENCE]
Problems I've watched them hit repeatedly: [LIST WHAT YOU'VE ACTUALLY SEEN]
For each problem: what form of paid resource would genuinely solve it, and
what form would look like a solution while leaving them stuck?
Rank by how much of the problem the resource actually removes.Product: [DESCRIPTION]
Audience: [AUDIENCE]
Price: [PRICE]
What must I be able to answer before launching? Group into: things I can
answer now, things I need to test, things I'm currently guessing at.
Flag anything where a wrong guess is expensive to reverse.What not to delegate
| Task | Why it fails |
|---|---|
| The opinion | A model has no stake. It will produce a defensible position, which is not the same as one worth reading. |
| First-hand detail | Anything it invents about your experience is fabricated by definition. This is the material that makes an article non-commodity. |
| Numbers and citations | Plausible-looking figures with no source are the most common failure. Verify every number or state it qualitatively. |
| Final voice pass | Model prose converges on a house style readers now recognise. The last edit should be yours. |
| What matters most | Prioritisation depends on your goals and constraints, which it can’t see. Ask for options, not decisions. |
Diagnostics
| Symptom | Cause | Fix |
|---|---|---|
| Output could apply to anyone | No audience or purpose supplied | Add who it’s for and what they’ll do with it |
| Confidently wrong facts | No permission to be uncertain | Add: “Where you’re not confident, say so rather than guessing” |
| Ignored half your instructions | Too many competing constraints | Split into two prompts — one job each |
| Right content, wrong format | Output shape unstated | Name the structure: table, ranked list, quoted passages |
| Rewrote when you wanted a review | Diagnostic intent unstated | Add: “Don’t rewrite. Report what you find.” |
| Copied your example too literally | Models attend closely to examples | Cut to one example, or describe intent instead |
| Bland despite a detailed prompt | Constraints describe form, not substance | Add a real claim, number, or observation of your own |
Questions
Yes. Task, context, constraints and output shape are model-independent. The differences are at the margins — some models accept XML-style tags to separate pasted source material from instructions, which helps on long inputs. Nothing above depends on a vendor-specific feature.
Put the stable parts there — your audience, your subject area, your voice preferences, what you never want it to do. Then the per-task prompt only carries what changes. This is the single highest-leverage change to a prompting workflow, because it removes the context you’d otherwise retype and forget.
Because a rewrite hands back model prose, and you then have to edit your voice back in. A diagnosis hands back a decision list, and you keep the writing. The revise stage is where models add the most value, and it’s the stage where accepting their output costs the most.
AI assistance isn’t penalised — mass-produced content with little value per page is, however it was made. The distinction Google draws is between content created to help a reader and content created primarily to rank. A pipeline like the one above, where the model diagnoses and you decide, sits on the right side of that line. Publishing forty auto-generated variations does not.
Next step
Take your last published article and run the generic-detection prompt from stage 5 on it. It’s the fastest way to see whether any of this is worth adopting — and the results are usually uncomfortable enough to be useful.