TL;DR: Fifty before and after prompts, grouped by the ten rewrites that do the work. Every pair shows only prompts, never outputs, because the teaching is in what the second version specifies that the first left the model to guess. Skim the move you keep getting wrong, and steal the wording.
Most before-and-after prompts published online pair a strawman nobody would type with a screenshot of a model output. Both halves fail: the strawman teaches nothing because you were never going to send it, and the screenshot cannot be checked or reproduced next month.
So this gallery is prompts only: fifty requests people genuinely send, each rewritten so the thing the model would otherwise guess is stated. The sections are organised by the fix, not the topic, which is the point: ten moves that transfer, not fifty templates that do not.
How Do You Read These Pairs?
Each pair is numbered, tagged with its job and with the move it demonstrates, then followed by one sentence naming what the rewrite specified. Read that sentence: the wording is disposable, the move is not.
Within each section, five pairs apply one move to five jobs: read down to learn the move, or across the table to match your week.
| If your job is | Jump to pairs | The move that usually fixes it |
|---|---|---|
| Email and client comms | 1, 6, 16, 22, 28, 31, 43 | Name the reader |
| Code review and debugging | 2, 11, 20, 23, 26, 35, 38, 48 | Bound the scope |
| Data and analysis | 3, 8, 12, 25, 27, 34, 37, 46 | Define what done means |
| Research and reading | 9, 15, 17, 30, 36, 44, 50 | Allow an honest I don't know |
| Planning and decisions | 4, 13, 18, 29, 40, 41, 47 | Split it with a checkpoint |
| Teaching and explaining | 7, 14, 24, 32, 39, 45 | Show one worked example |
| Creative and marketing copy | 5, 10, 19, 21, 33, 42, 49 | Say what you want, not what you don't |
For ten of these written out at length, see 10 bad prompts, fixed. This page is the browsable version: more pairs, less commentary.
How Do You Turn a Vague Ask Into a Specified Deliverable?
By naming the artefact, its format and its length: three guesses the model makes silently when you leave them out. A vague ask has no shape, so anything shaped comes back.
1 · Email · specify the deliverable
BEFORE
Write something to the client about the delay.
AFTER
Write a client email about the two-week delay to the design handoff.
Four sentences, under 150 words: what slipped, the new date, what we
are doing about it, one ask. Plain text, no subject line.
Names the artefact, the length and the four required parts, so none is left open.
2 · Code review · specify the deliverable
BEFORE
Review this PR.
AFTER
Review this PR. Return at most 8 findings as a markdown list. Each
finding: file and line, severity (blocker / should-fix / nit), one
sentence on the problem, one on the fix. No summary paragraph.
[PASTE DIFF]
Caps the count, fixes the severity vocabulary and bans the summary, which makes the review triageable.
3 · Analysis · specify the deliverable
BEFORE
Analyse this sales data.
AFTER
From the CSV below, produce a markdown table of the 10 accounts with
the largest revenue change quarter over quarter. Columns: Account |
Q2 | Q3 | Change % | one-line note. Then three bullets on what moved.
Nothing else.
[PASTE CSV]
Analyse names an activity, not an output. The rewrite names the table, the columns and the stopping point.
4 · Planning · specify the deliverable
BEFORE
Help me plan the launch.
AFTER
Draft a four-week launch plan for a paid Chrome extension, as a table
with columns Week | Workstream | Owner role | Deliverable | Done when.
Week 4 ends on public launch day. Maximum 20 rows.
A plan can be a paragraph, a checklist or a Gantt chart. The columns pick one.
5 · Creative · specify the deliverable
BEFORE
Write some social copy for the new feature.
AFTER
Write 5 LinkedIn post options announcing shared team prompt libraries.
Each: 40-60 words, one concrete before-and-after scenario, no hashtags,
no emoji, first line under 12 words so it survives the see-more cut.
The see-more constraint encodes a platform fact the model cannot read off your intent.
If you are not sure which container you want, how to decide what output format you actually need is the step before this one.
Who Is the Reader, and What Do They Already Know?
Every prompt has one intended reader whether or not you name them. Leaving the slot empty produces an answer aimed at an average reader who does not exist.
6 · Email · name the reader
BEFORE
Explain the pricing change to customers.
AFTER
Write an email explaining the pricing change to existing annual
customers who are mid-term and unaffected until renewal. They know our
plan names; they have never seen our internal tier codes. Open by
saying nothing changes today.
Names who is reading, what they already know, and the sentence that matters most to them.
7 · Teaching · name the reader
BEFORE
Explain recursion.
AFTER
Explain recursion to a bootcamp student in week 3 who is comfortable
with loops and functions and has never seen a call stack. One everyday
analogy, then one Python example under 12 lines, then the base-case
mistake beginners make.
Prior knowledge decides whether an explanation lands as patronising or unusable, and it is invisible unless typed.
8 · Analysis · name the reader
BEFORE
Summarise these support tickets.
AFTER
Summarise these tickets for the on-call engineer starting a shift in
ten minutes. They know the system; they do not know what happened
overnight. Rank by what is still broken, not by ticket volume.
[PASTE TICKETS]
The reader and their deadline decide the ordering, where most of a summary's value sits.
9 · Research · name the reader
BEFORE
Summarise this paper.
AFTER
Summarise this paper for a policy analyst with no statistics training
who briefs a minister tomorrow. Say what the finding supports and what
it does not support. No formulae, no methods section.
[PASTE PAPER]
A statistician and a policy analyst need opposite summaries of one paper. Only one is in your head.
10 · Creative · name the reader
BEFORE
Write a landing page headline.
AFTER
Write 8 landing-page headlines for a solo freelancer who already uses
ChatGPT daily and is sceptical of anything marketed as an AI tool.
They know what a prompt is. Do not explain AI. Maximum 9 words each.
Naming the reader's existing scepticism rules out the entire category of copy that explains the obvious.
What Would Make This a Good Answer?
Write down the test the answer has to pass. A criterion you are holding in your head is not available to the model.
11 · Code review · define what done means
BEFORE
Make this function better.
AFTER
Refactor this function. A good answer keeps the signature and observable
behaviour identical, removes the nested loop, and stays under 40 lines.
If you cannot do all three, say which one you dropped and why.
Three checkable conditions plus an escape hatch, so the reply is either compliant or explicitly non-compliant.
12 · Analysis · define what done means
BEFORE
Find insights in this data.
AFTER
From this data, report only findings that hold across all three months,
involve at least 50 accounts, and could be acted on this quarter.
Anything failing one of those three, leave out.
[PASTE DATA]
"Insight" is undefined. Persistence, sample size and actionability are not, and each one is a filter.
13 · Planning · define what done means
BEFORE
Write our Q4 OKRs.
AFTER
Draft 3 objectives for Q4. A usable objective here is one sentence,
has a number in its key result, and would be settled by a dashboard we
already have. Name the dashboard for each one.
Tying each objective to an existing dashboard quietly rules out the aspirational filler this task attracts.
14 · Teaching · define what done means
BEFORE
Make a quiz on this chapter.
AFTER
Write 10 questions on this chapter. A good question cannot be answered
by pattern-matching the headings, has exactly one defensible answer,
and takes under 90 seconds. For each, name the sentence it tests.
[PASTE CHAPTER]
Question quality is otherwise subjective. These three conditions make it inspectable in seconds.
15 · Research · define what done means
BEFORE
Find sources on remote work and productivity.
AFTER
List candidate sources on remote work and productivity. Include one
only if you can name its study design (RCT, panel, survey, case study)
and its sample size. If you cannot name both, leave it out and say so.
An inclusion rule doubles as a hallucination filter, being harder to satisfy than a bare list.
Where Should the Actual Ask Sit in a Prompt?
First. A request buried inside context and pleasantries competes with everything around it for the role of the task. Put the ask on line one and let the material follow.
16 · Email · put the ask first
BEFORE
Hey, so I was thinking about what we discussed last week and I know
you're slammed but if you get a chance could you maybe look at whether
the onboarding emails are any good? Anyway, hope the launch went well.
AFTER
TASK: Rewrite the 3-email onboarding sequence below.
CONSTRAINTS: each under 120 words, keep every CTA link text as-is,
do not change email 1's subject line.
CONTEXT: sent to trial users on days 0, 2 and 6.
SEQUENCE: [PASTE]
Shorter than the original and unambiguous about which sentence is the instruction.
17 · Research · put the ask first
BEFORE
[Three paragraphs of background about the project, the team and the
timeline] ... so what I want to know is whether any of this is
actually supported by the notes.
AFTER
QUESTION: Which of the three claims below is supported by the attached
notes, and which is not? Answer claim by claim.
CLAIMS: 1) ... 2) ... 3) ...
NOTES: [PASTE]
Separating question from evidence forces you to state the claims as claims, half the work already.
18 · Planning · put the ask first
BEFORE
We've had a rough quarter, churn is up, the team is tired, marketing
wants more content, support is drowning and I need to work out what to
do about headcount.
AFTER
DECISION: where to put one new hire next quarter.
OPTIONS: support, content, or platform engineering.
EVIDENCE: [3-5 bullets of the numbers above]
OUTPUT: one recommendation, the two strongest arguments against it,
and what would change your mind.
Turns a vent into a decision with named options, the difference between sympathy and analysis.
19 · Creative · put the ask first
BEFORE
[Long brand story: founding, values, the customers we love, the tone
we're going for, the competitors we don't want to sound like]
Can you write a tagline?
AFTER
DELIVERABLE: 10 taglines, maximum 6 words each.
BRAND NOTES (reference only, do not quote or paraphrase these lines):
[PASTE STORY]
Marking the brief as reference stops it being read as the thing to rewrite.
20 · Code review · put the ask first
BEFORE
[400-line file pasted]
Thoughts?
AFTER
ASK: find why /api/export returns 500 for users with more than 1,000
rows.
SCOPE: only the handler and the query builder below.
CODE: [PASTE HANDLER AND QUERY BUILDER]
A specific question against 40 lines is a different request from an open invitation against 400.
What Do You Want, Rather Than What You Do Not Want?
A ban names one region to avoid and leaves everything else open, which is why don't be formal and don't use jargon do so little work. Say what to do instead, then keep the one narrow ban that matters.
21 · Creative · say what you want
BEFORE
Don't make it sound like AI wrote it.
AFTER
Write in short declarative sentences, averaging under 18 words. Use we
and you. No adjective may appear twice in the same paragraph. Never
open a sentence with the phrase In today's.
Three positive constraints and one narrow, checkable ban, rather than one instruction about a vibe.
22 · Email · say what you want
BEFORE
Don't be too formal.
AFTER
Write it the way I would speak to a colleague I like: contractions,
first names, one idea per sentence, no throat-clearing opener. Sign off
with Thanks, and my first name.
Not formal spans friendly through flippant. The rewrite picks a register and names its markers.
23 · Code review · say what you want
BEFORE
Don't suggest big rewrites.
AFTER
Restrict suggestions to changes under 20 lines touching a single file.
If the real fix is larger, describe it in one sentence under a heading
Out of scope, and stop there.
Puts a number on big, and gives the oversized suggestion somewhere legitimate to go.
24 · Teaching · say what you want
BEFORE
Don't use jargon.
AFTER
Define every technical term in the sentence where it first appears, in
under 12 words. If a term cannot be defined that briefly, use the
everyday phrase and put the technical term in brackets after it.
A rule you apply term by term, not a judgement call about what counts as jargon.
25 · Analysis · say what you want
BEFORE
Don't make anything up.
AFTER
Every number in your answer must be traceable to a cell in the table
below. After each figure, put its column and row label in square
brackets. If a figure is not in the table, write not in data rather
than estimating.
Traceability is a mechanism; the original was a wish. The permission section below is its real fix.
Negative instructions have their own page: why don't do X often backfires in prompts covers what vendors document about precedence and conflict, with thirty more rewrites.
Which Piece of Context Actually Changes the Answer?
The one you know and never typed. Context earns its place when the answer would be different without it, which is stricter than might be relevant.
26 · Code review · add the deciding context
BEFORE
Why is this query slow?
AFTER
Why is this query slow? Postgres 17, the table has 40M rows,
created_at is indexed and status is not, it runs hourly from cron, and
it got slow the week we added the status filter.
[PASTE QUERY]
Six facts that change the diagnosis, none of them visible in the SQL itself.
27 · Analysis · add the deciding context
BEFORE
Is this conversion rate good?
AFTER
Is this conversion rate good? B2B SaaS at $4.99/mo self-serve, 80% of
traffic from a Chrome Web Store listing, conversion here means install
to paid within 30 days, and we changed onboarding two weeks into the
window.
[PASTE NUMBERS]
"Good" is meaningless without the business model, the traffic source and the definition being measured.
28 · Email · add the deciding context
BEFORE
Write a follow-up to this prospect.
AFTER
Write a follow-up. What matters: we met at a conference, they asked one
specific question about SOC 2 that I could not answer at the time, I
can now (report available under NDA), and they said to ask again in
September. It is late August.
The unanswered question is the entire reason this email has a reason to exist.
29 · Planning · add the deciding context
BEFORE
Should we build this feature?
AFTER
Should we build this? Two engineers, six weeks, and those same two
engineers carry the pager. Requested by 4 of our 12 largest accounts,
none of whom have threatened to leave. We ship weekly.
[PASTE FEATURE SPEC]
Capacity, on-call load and customer pressure decide this, and none of them are in the spec.
30 · Research · add the deciding context
BEFORE
What should I read about prompt engineering?
AFTER
What should I read about prompting? I have already read the OpenAI and
Anthropic docs and want primary sources: papers, vendor changelogs or
official docs. I care about instruction-following, not architecture.
English only, last 18 months.
Saying what you have already read removes the introductory tier a bare question invites.
Can You Show One Example Instead of Describing the Pattern?
Yes, and one real example is usually shorter than the paragraph describing it. This is few-shot prompting at its smallest useful size: one sample fixing shape, length and voice at once.
31 · Email · show one worked example
BEFORE
Write replies in my voice.
AFTER
Match the voice of this reply I sent last week:
[PASTE ONE REAL 80-WORD EMAIL]
Now reply to the message below in the same voice: same greeting style,
same sign-off, same rough length.
Voice is easier to demonstrate than describe, and a real email carries details you would not think to specify.
32 · Teaching · show one worked example
BEFORE
Write flashcards from this chapter.
AFTER
Write flashcards in exactly this shape:
Q: What does a base case do in recursion?
A: Stops the recursion. Without one, calls repeat until the stack runs out.
(One sentence of function, one of consequence.) Produce 15 from the
chapter below.
The example fixes the card shape and the level of detail at once.
33 · Creative · show one worked example
BEFORE
Write product descriptions like ours.
AFTER
Here is one of ours we like:
[PASTE 45-WORD DESCRIPTION]
The pattern: one use-case sentence, one materials sentence, one care
sentence, no adjectives in the first line. Write 6 more in that pattern
for the products below.
Naming the pattern turns one sample into a rule, not something to imitate loosely.
34 · Analysis · show one worked example
BEFORE
Categorise these support tickets.
AFTER
Categorise these tickets using only these labels: billing, bug, how-to,
feature-request, abuse. Two worked examples:
"Card declined but I was charged twice" -> billing
"How do I export my library?" -> how-to
Output CSV with two columns: ticket_id,label.
Two examples pin down the boundary cases that a label list on its own leaves ambiguous.
35 · Code review · show one worked example
BEFORE
Write tests for this module.
AFTER
Write tests in the style of this existing test:
[PASTE ONE 15-LINE TEST]
Same framework, same naming convention, same fixture helper. Cover the
three branches in the function below.
House conventions live in the code, not your description of it. Paste one test and they come along.
Have You Given the Model Permission to Say I Don't Know?
Not unless you say so. Absent permission, the shape of a helpful reply is a confident one, so name the honest fallback and make it an acceptable answer.
36 · Research · allow an honest I don't know
BEFORE
What does the research say about X?
AFTER
What does the research say about X? Where you are confident, name the
study and year. Where you are not, write uncertain and say what would
settle it. The literature does not settle this is a complete answer.
Makes the honest answer explicitly acceptable, which stops it competing with a confident one.
37 · Analysis · allow an honest I don't know
BEFORE
What caused the drop in signups?
AFTER
Give the two most likely causes of the drop given the data below, then
the single check that would distinguish them. If this data cannot tell
them apart, say so instead of picking one.
[PASTE DATA]
Asking for a discriminating check rather than a verdict turns a guess into something actionable.
38 · Code review · allow an honest I don't know
BEFORE
Is this code secure?
AFTER
Review this handler for injection, authorisation and secret-handling
issues. Anything you cannot judge without seeing the caller or the
schema, list under a heading Need to see, rather than assuming.
[PASTE HANDLER]
Security review depends on code outside the snippet, so name where those gaps get recorded.
39 · Teaching · allow an honest I don't know
BEFORE
Explain what this error means.
AFTER
Explain this error. If more than one cause produces it, list them all
with the tell that distinguishes each. If the traceback alone is not
enough to decide, say what extra output you need.
[PASTE TRACEBACK]
Most error messages have several causes, so ask for the distinguishing tell rather than the cause.
40 · Planning · allow an honest I don't know
BEFORE
Estimate how long this will take.
AFTER
Estimate this work in engineer-days as a range. Name the assumption
that drives the width of the range, and what you would need to know to
halve it. If part of the spec is too vague to estimate, say which part.
An estimate with its uncertainty named is usable. A single number hiding it is not.
What Is Explicitly Out of Scope?
Whatever you would reject on sight. Exclusions cost one line and remove entire branches the request would otherwise spend effort on, which makes them the strongest move on open-ended asks.
41 · Planning · bound the scope
BEFORE
Give me ideas to grow the business.
AFTER
Give me 5 growth ideas we could start this month with no new hires, no
paid ads, and no engineering beyond one day of work. Exclude anything
requiring a partnership, a conference or a rebrand.
Four exclusions rule out most of what this question usually returns, and they take one sentence.
42 · Creative · bound the scope
BEFORE
Redesign our onboarding.
AFTER
Propose changes to onboarding that touch only the copy and the ordering
of the 4 existing screens. No new screens, no new API calls, no video.
Maximum 6 changes, each one sentence.
Naming what cannot change keeps the proposals inside what you can ship this week.
43 · Email · bound the scope
BEFORE
Write a cold email.
AFTER
Write a cold email under 90 words to a named recipient. No links, no
attachments, no calendar link, exactly one question, and nothing that
would be untrue if the recipient has never heard of us.
The last clause excludes the false familiarity that makes cold email read as spam.
44 · Research · bound the scope
BEFORE
Research our competitors.
AFTER
Cover these four named competitors only: A, B, C, D. Use publicly
documented facts from their own pages. Exclude pricing, exclude review
scores, exclude anything you would source from an aggregator blog.
Excluding the fast-moving and badly-sourced fields is what makes the result worth keeping.
45 · Teaching · bound the scope
BEFORE
Teach me SQL.
AFTER
Teach me exactly the SQL needed to answer questions about a single
table: SELECT, WHERE, GROUP BY, ORDER BY, LIMIT. No joins, no
subqueries, no window functions. Then four exercises on the schema below.
A curriculum bounded by what you need this week is finishable. "Teach me SQL" is not.
Should This Be One Prompt or a Sequence With a Checkpoint?
Split it whenever a wrong assumption early would waste the whole answer. A checkpoint costs one turn, which is cheap next to two thousand words built on a misunderstanding.
46 · Analysis · split it with a checkpoint
BEFORE
Analyse this dataset and write the report.
AFTER
Step 1: list what you think each column contains, and any rows you
cannot interpret. Stop there and wait for my confirmation before
analysing anything.
[PASTE DATA]
Column semantics are the assumption everything downstream rests on, so that is where the checkpoint belongs.
47 · Planning · split it with a checkpoint
BEFORE
Write the PRD.
AFTER
Before drafting, ask me up to 6 questions whose answers would change
the PRD. Wait for my answers. Then draft only the Problem and
Non-goals sections and stop for review.
Questions first is a general-purpose move, covered in how to get ChatGPT to ask you questions first.
48 · Code review · split it with a checkpoint
BEFORE
Migrate this service from X to Y.
AFTER
Step 1: list the files that must change and the order to change them,
one line of reasoning each. Do not write any code yet. When I reply go,
do file 1 only and stop.
The plan is reviewable in a minute. Three files built on a wrong one are not.
49 · Creative · split it with a checkpoint
BEFORE
Write the launch blog post.
AFTER
Step 1: give me 3 possible angles, one sentence each.
Step 2 (after I choose): an outline, H2s only.
Step 3 (after I approve the outline): the draft.
Do step 1 now and nothing else.
Angle determines everything else, and it is the cheapest thing to get wrong early.
50 · Research · split it with a checkpoint
BEFORE
Write a literature review on X.
AFTER
Step 1: propose the inclusion and exclusion criteria, and the 5 themes
you would organise the review around. Wait for my edits. Then draft
section by section, one section per reply.
Criteria and structure are where reviews go wrong, and both are cheapest to fix before the prose exists.
Does a Longer Prompt Beat a Shorter One?
No, and this page is not an argument that it does. The claim is that each after is more specified than its before, which is a different property from longer. Every rewrite here removes a guess; padding adds words without removing any.
The five rewrites in the put-the-ask-first section are shorter than the prompts they replace, because the problem there was a buried instruction, not a missing detail. The full argument, with vendor sources, is in does a longer prompt get better results. Read it before concluding that the lesson of fifty pairs is write more.
Nothing here promises a better answer either. Specification removes guesswork, which is a precondition for a good answer rather than a guarantee of one. A perfectly specified prompt asking for the impossible still will not work.
What Does Prompt Architects Do With This?
Roughly this transformation, automatically. The Enhance button in our extension restructures a casual prompt into one carrying a role, explicit instructions, constraints and an output format: the same class of change as the pairs above. Our FAQ page publishes the free plan's limit as "5 prompt enhancements per day, forever", with no card required, verified live on 29 August 2026. The /pricing page shows no Free price card, though its own FAQ says "The Free plan is free forever", so that daily figure comes from the FAQ page.
The honest part: you do not need us for any of this. Every move here is something you can type yourself in four seconds once you know which one is missing, and the people who get most from a tool like ours already know that. What it buys is consistency on the fiftieth prompt of the day, not a technique we invented.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An Account