AI Workflow Before After Examples: 10 Real Prompt Rewrites

Why one-line prompts fail at work

A one-line prompt is not wrong, it is under-specified. The model has to guess your role, your reader, your format, and which facts count — and it guesses the average of everything it has ever seen. The average of every meeting summary on the internet is a neat recap that assigns nothing to nobody.

That is also why the same prompt gives you a different answer on Tuesday. You never pinned down the variables, so they float. If you have hit that already, the AI Execution Loop is the wider system this article sits inside — Think, Structure, Execute, Refine. What follows is the Structure step, done ten times on real work.

Most professionals reach for a longer prompt when the answer disappoints. Length is not the fix. A 200-word request that still never names the reader produces the same fog, slower.

There is a second cost nobody counts. The bad answer is not free — you read it, judge it, and then spend ten minutes rewriting it into the thing you wanted. Across a week that is most of an afternoon spent editing drafts you should never have received.

The Four-Line Rewrite, explained

The Four-Line Rewrite is the approach we teach at PromptPeakAI for turning an ask you already typed into a brief the model can work from: rewrite it as four lines — who the AI is, what the finished thing is, which facts it may use, and what shape the answer takes. It is the same four ideas as Role, Goal, Context and Format, applied as a repair job rather than a blank page.

Write the job title of the person whose desk this work normally lands on. “You are a senior operations manager” changes vocabulary, assumed knowledge, and what the answer bothers to explain. Seniority matters more than industry — a director writes shorter than an analyst. There is more on this in the guide to writing a role that lands.

“Help with the report” is a topic. “A one-page status update for my director, covering progress, risks and asks” is a deliverable. If you cannot picture the document at the end of the sentence, the model cannot either.

This is the line almost everyone skips, and it is the one that stops invented numbers. Say where the facts come from and what to do when they run out: use only the text below, and write NOT STATED for anything missing. That single instruction is the difference between a draft you check and a draft you cannot trust. It works because it gives the model a legal answer for a gap — see negative constraints for why banning things outperforms asking nicely.

Columns, word count, number of sections, tone. “A four-column table: task, owner, deadline, dependency” removes every layout decision from the model. The format line is what makes output pasteable instead of re-typeable.

You are a [job title, seniority level].

Produce: [the finished deliverable, in one phrase].

Use only the [transcript / email / table] pasted below. If a detail is missing, write NOT STATED. Do not estimate, assume, or add anything that is not in the text.

Format: [exact shape — columns, sections, word count, tone].

---
[paste your source text here]

AI workflow before after examples: meetings and reports

These first three AI workflow before after examples cover the tasks that eat Monday mornings. In each pair, the before is a real thing people type and the after is the same request with the four lines added. Nothing else changed — same tool, same source document, same day.

Example 1: the meeting transcript

Forty minutes of recording, half of it small talk, and you need to know who owes what by Friday.

The fenced facts line is doing the work here. Without it the model helpfully assigns owners based on who talked most. The full walkthrough lives in the guide to turning a transcript into minutes.

Example 2: the weekly status

Your director reads eleven of these before lunch. Yours needs to be skimmable in twenty seconds.

“Writing upward to a director” is the whole trick. Naming the reader, not just the writer, is what strips the filler out.

Example 3: the executive summary

A sixty-page document landed and someone wants the gist by 2pm. This one is better as a template you keep.

You are a chief of staff summarising for a board that will not read the source.

Produce: a five-bullet executive summary, plus one line naming the single decision this document is asking for.

Use only the document below. Where the document is vague, say so explicitly rather than resolving it. Write NOT STATED for anything absent.

Format: five bullets, maximum 25 words each, plain language, no adjectives. Then one line beginning "Decision required:".

Examples 4 to 7: money, vendors, scope

This is where the fenced-facts line stops being good practice and starts being self-defence. Every task below involves numbers somebody will act on.

Example 4: the budget variance

Actuals versus budget, and a non-finance audience who needs the story, not the spreadsheet.

“Do not infer causes” is the line that saves you. Plausible-sounding reasons for a spike are exactly what a language model is best at inventing. The longer version is in writing a financial narrative for leadership.

Example 5: the vendor shortlist

Three proposals, twenty pages each, and a comparison that has to survive being questioned.

You are a procurement lead preparing a shortlist paper for a decision meeting.

Produce: a comparison matrix of the three proposals attached.

Assess only on: price, implementation timeline, support commitment, security certifications. Use only what each proposal states in writing. Where a proposal does not address a criterion, write NOT STATED — do not score it, and do not fill the gap from the vendor's website.

Format: one row per vendor, one column per criterion, plus a final column listing what is missing from that proposal.

Example 6: the scope audit

The client keeps adding “quick” requests by email and nobody has checked them against the signed scope.

Asking for the quoted clause is a verification trick worth stealing everywhere: it forces the model to point at the evidence instead of asserting a conclusion.

Example 7: the RFP answer

Sixty questions, most of which your team has answered before in slightly different words.

You are a bid manager answering a public-sector RFP.

Produce: a draft answer to the question below, using our previous answers pasted underneath as the only source of fact.

Do not invent capabilities, certifications, client names, or numbers. If our previous answers do not cover part of the question, write GAP — NEEDS INPUT and name exactly what is missing.

Format: 150 words maximum, second person, no marketing adjectives, no superlatives.

That GAP instruction turns the draft into a to-do list for the humans. More on the whole cycle in the guide to writing RFP responses.

Examples 8 to 10: people and clients

The last three carry reputational weight. A wrong number in a variance note is embarrassing; a wrong sentence in a review is a formal problem.

Example 8: the performance review

Six people, a deadline, and a pile of notes you took across the year.

You are a line manager writing a mid-year review for a direct report.

Produce: a draft review with three strengths and two development areas.

Use only the notes and feedback pasted below. Every point must cite a specific incident from those notes. If a competency has no evidence in the notes, write NO EVIDENCE RECORDED rather than writing something general.

Format: for each point — the behaviour, the example, the impact. No personality descriptions, no comparisons to other team members.

“No personality descriptions” is a negative constraint with a legal reason behind it. Reviews describe behaviour and impact, not character. The manager-side detail is in writing performance reviews fairly.

Example 9: the sales follow-up

A good discovery call, and a follow-up email that usually reads like it was sent to everyone.

Example 10: the client delay notice

You are late. The email has to say so without either grovelling or hiding it.

“Apologise once and not again” is the sort of instruction you only write after reading an AI draft that apologised five times. Lay the ten AI workflow before after examples side by side and one line does most of the work in each — usually a different one:

When the rewrite still goes wrong

The four lines fix the common failures, not all of them. Three things go wrong often enough to be worth naming.

That third row is worth understanding rather than memorising. Anthropic’s engineering write-up on effective context engineering for AI agents, published in September 2025, describes the effect as context rot: as the number of tokens in the window grows, the model’s ability to accurately recall information from that window drops. It is not the model ignoring you. It is attention thinning out across a long conversation.

The practical consequence is small and slightly annoying: on turn four or five, re-anchor. One short block does it.

Reapplying my rules from the start of this chat before you answer:

- Role: [paste your role line again]
- Source: use only the text I pasted. Missing details are NOT STATED.
- Format: [paste your shape line again]

Now: [your next request]

If the answer is close but not right, resist the urge to start again. Change one line, re-run, and see what moved — the method for that is in refining output without starting over.

What the rewrite cannot fix

Four lines make the output usable. They do not make it true, and they do not make it yours. Three things stay on your desk permanently.

Most professionals overcorrect once, hand something over that needed a human, and then settle into the right split. It is worth reaching that point deliberately rather than by accident. If you want the wider verification habit, the guide to fact-checking AI research covers the checking pass properly, and why AI invents things covers the underlying mechanism.

Which plan you actually need

None of this requires a subscription. Every rewrite above is plain text into a chat box, which the free tier of every mainstream assistant handles. Paid tiers raise upload and length limits; they do not unlock the method.

The real constraint is not price, it is what you are allowed to paste. That depends on your account type, not your plan tier. OpenAI’s help centre for ChatGPT account and data questions states the boundary plainly: content from business offerings is not used to train models, while content from individual accounts may be unless you opt out — and specific prompts cannot be deleted from your history afterwards. Other vendors draw the line in the same place.

Redaction is easier than it sounds: swap identities for roles before you paste, and the output comes back in the same terms.

• Client and vendor names become Client A, Vendor B.
• Colleagues become the finance director, the project lead.
• Real figures become [FIGURE] when the number itself does not change the writing.

Pick one task from the ten above — ideally the one you will do again this week — and rewrite it four lines at a time. The second one takes a minute. By the fifth, you stop thinking about the structure and just write briefs instead of asks.

Then keep them. The genuine compounding effect of AI workflow before after examples is not the ten on this page, it is the four or five you write yourself for the tasks that actually recur on your desk. A saved note with your own role line and your own format line, ready to paste, is worth more than any prompt list — because it already knows who you are and who you write for.