Prompt library — ready to copy, paste, and use

It's 11 PM on a Sunday. You just wrote a prompt that worked — a work email came out clean, without that corporate tone you were trying to avoid. You copy it into a note on your phone. "I'll use this again tomorrow," you think.

Tuesday you need it. You open the notes app. It's not there. You'd saved it in the other one, on your laptop. Or in the WhatsApp thread you send to yourself. Or in an untitled Google Doc.

That prompt that worked, you write again from scratch. It takes fifteen minutes instead of five. And this time, the result doesn't quite land.

The cost of not having a library isn't that you lose prompts. It's that you think the same problem twice.

Five prompts worth saving this week

Here are five, ready to copy. Each one has the four CAFÉ letters labeled inside in brackets — so you see how they're built while you use them.

1. Professional email (the one you send every week)

`` You're an independent consultant writing to someone you barely know. [C] Context: I'm writing to propose a 20-minute meeting about a specific project. [A] Action: write the email. [F] Format: maximum 10 lines, include subject line, no corporate jargon, one single question at the end. [E] Example of the tone I like: "Hi Laura, I'm writing because I saw your talk at the festival and the idea of X stuck with me..." ``

What each part does: [C] gives role and situation. [A] states the task in one verb. [F] sets the hard constraints — lines, tone, structure. [E] shows how it sounds when you write.

2. Meeting recap (from a long transcript)

`` [C] This is the transcript of a 50-minute meeting with my team about a product launch. [A] Pull three things: decisions made, open action items with owner, and topics left unresolved. [F] Format: three blocks with headers, short bullets, no paraphrasing or embellishment. [E] If something wasn't clear in the meeting, mark it "[ambiguous]" instead of inventing it. Transcript: [paste here] ``

What each part does: [E] here isn't an output example, it's a behavior rule — "if you don't know, say so." That kills about 80% of the hallucinations you'd otherwise have to review.

3. LinkedIn post (about something you learned)

`` [C] I work as [profession] and I just learned something on a real project: [describe in two lines what it was]. [A] Write a LinkedIn post in first person telling the experience. [F] Max 150 words, opens with a short sentence that isn't a cliché ("today I learned" is banned), one idea per paragraph, closes with a real question. [E] Tone: the tone of someone telling a story over coffee, not someone giving a TED talk. ``

What each part does: Banning a specific cliché ("today I learned" is banned) works better than asking for creativity in the abstract. The model doesn't know what a cliché is until you name it.

4. Comparison table (for deciding fast)

`` [C] I'm evaluating three options: [A], [B], [C]. My main criterion is [price/speed/whatever]. [A] Build a comparison table. [F] Columns: option, main pro, main con, price, recommendation in one line. [E] In the recommendation row don't say "it depends" — pick the one the data favors for my criterion and justify it in six words. ``

What each part does: Forcing the recommendation in [E] avoids the polite AI answer that says nothing. If you don't ban "it depends," it'll say it.

5. Social media image (descriptive prompt)

`` [C] I need a vertical image for Instagram to go with a post about remote work. [A] Generate a photorealistic image. [F] Vertical 4:5 composition, medium shot, natural window light, warm palette. [E] Scene: a wooden desk with a steaming coffee cup next to an open laptop, a blurred plant in the background. No faces, no text on the image, no logos. ``

What each part does: [E] here is basically the whole image. Without a scene described in physical detail, the model gives you the stock photo it has closest. With a scene, it gives you something that looks thought-through.

What to take away

1. Save every prompt you've used twice. Twice is the threshold. If you wrote it twice, you'll write it ten times — no reason to.

1. Adapt, don't blindly copy. A prompt without your context in it is generic. Swap names, examples, tone. The skeleton belongs to the prompt; the flesh is yours.

1. If CAFÉ isn't there, the prompt limps. If the answer comes out weird, open the prompt and find which of the four letters is missing. Almost always it's Format or Example.

On a Tuesday in 2025, Lucas — a freelance copywriter in Buenos Aires, seven years into the trade — wrote his last corporate email from scratch. It took him forty minutes. Three rewrites, two coffees, a moment of "just send it, it's fine." The next day he had another one like it.

That evening he sat down to do something he'd been putting off: turn that email into a saved prompt. He pulled out the client data, left the variables empty, wrote a line at the top of the file with what worked and what didn't. Fifteen minutes.

The next week the same type of email took him seven minutes: five to fill the variables, two to review the output. The point isn't that he saved time. The point is he stopped thinking the same problem twice.

That's a prompt library. Not a collection of hits. A set of decisions you've already made — about tone, format, examples, constraints — so you don't have to make them again every Monday.

What goes in the library and what doesn't

Hard rule: if you'll do it once in your life, don't turn it into a saved prompt. Write it, send it, forget it. The library is for what repeats.

What repeats in real work: emails by type (sales, collections, follow-up, intro), synthesis by format (meetings, long documents, customer feedback), social content (LinkedIn posts, newsletters, images), recurring structures (dashboards, comparison tables, briefs), and personal automations (editing your own writing, translating in your voice, sorting loose notes).

Ten prompts cover most of the weekly volume of an average professional. You don't need fifty. You need ten that work and that you can find fast.

The ten prompts, with adaptation notes

Professional writing — 3 prompts

1. Soft sales email (asks for a conversation, not a close)

“ You're [role], writing to [decision-maker function] in [industry]. [C] Context: they didn't contact you; you spotted the opportunity. [A] Write an email proposing a 20-minute conversation. [F] Max 120 words, include subject, single CTA, no marketing adjectives. [E] Tone: peers, not vendor. Details: prospect [name], company [X], problem I identified [Z], my specific solution [W]. “

How to adapt: change role and industry. The critical part is the final [E] — if you don't name the specific problem you spotted, the AI writes something generic that sounds like a newsletter.

2. Post-proposal follow-up (no begging)

“ [C] I sent a proposal [X days] ago to [name] for [service]. No reply. [A] Write a short follow-up. [F] Max 6 lines, confident without being pushy, offering help, with link to the proposal. [E] Tone: as if you knew your proposal is worth it — because it is. No "sorry to bother." “

How to adapt: this prompt works if in [E] you explicitly ban submissive phrases. If you don't, the AI defaults to mild groveling.

3. LinkedIn post with a real case

“ [C] I just finished a project with [client] in [industry]. Problem they had: [X]. Solution: [Y]. Measurable result: [Z with number]. [A] Write a LinkedIn post. [F] Max 180 words, opens with a line that isn't guru bait, three short paragraphs, closes with a real question. [E] Banned: "here's a story," "plot twist," "hot take," opening emoji. Tone: someone telling a story over coffee. “

How to adapt: the banned-phrases list is personal — add the ones that make you roll your eyes when you scroll LinkedIn.

Formats — 3 prompts

4. Comparison table with forced recommendation

“ [C] I'm evaluating [N options] with main criterion [X]. Data: [paste options with attributes]. [A] Build comparison table. [F] Columns: option, main pro, main con, [key metric], recommendation in one line. [E] In the recommendation row don't say "it depends" or "both make sense." Pick one and justify in six words. “

5. KPI dashboard ready to paste into Sheets

“ [C] Business: [type]. Scale: [N clients/products]. Metrics that matter: [list]. [A] Design dashboard structure. [F] Two tables: overview (one row per segment) and detail (one table per segment). Fixed columns: identity, target, actual, variance %, trend. [E] Output in markdown or CSV, ready to paste into Google Sheets. No commentary. “

6. Executive summary of a technical document

“ [C] I have a technical document of [X pages] on [topic]. Summary audience: [non-technical execs / board / investor]. Decision they need to make: [go/no-go / invest / prioritize]. [A] Produce executive summary. [F] Max 300 words, structure: what we saw → why it matters → 3 key findings → recommendation. [E] Zero jargon. If an unavoidable technical term appears, define it in six words. Document: [paste here] “

Images and social — 2 prompts

7. Editorial image for a blog post (vertical 3:4)

“ [C] I need an image for an article on [topic]. [A] Generate realistic photograph. [F] Vertical 3:4 composition, 1200x1600, natural documentary light, [warm/cool/contrast] palette. [E] Scene: [describe main object, surroundings, angle, the detail that anchors attention]. No identifiable faces, no text on the image, no logos. Editorial documentary style. “

How to adapt: the [E] part is 80% of the result. If you don't describe a scene with concrete objects, you'll get the stock photo the model has closest at hand.

8. Short social thread (Instagram/X)

“ [C] Topic: [X]. Destination format: [Instagram carousel / X thread / Threads]. [A] Write [N] chained pieces. [F] First piece: short hook without hashtags. Pieces 2 to N-1: one idea per piece, no closing. Last piece: simple CTA (question, link, or nothing). Length per piece: [platform limit]. [E] Banned: "🧵", "1/", "hot take," "unpopular opinion." “

Small apps and personal automations — 2 prompts

9. Text editor (your invisible hand)

“ [C] I'm a [role], I write [type of text]. [A] Edit the text below. [F] Tasks: fix spelling and grammar, shorten sentences over 25 words, swap fancy words for plain ones, keep the tone [formal/ casual/technical]. [E] Don't change structure or add ideas. If a sentence reads ambiguous, leave it and mark "[review]" in the margin. Return the edited text first; then a short list of important changes. Text: [paste here] “

10. Notes organizer (brain dump to structure)

“ [C] Below is a brain dump — loose notes on [topic], out of order, some half-written. [A] Organize them. [F] Output: three blocks — main ideas (three max), concrete tasks (verb-first), open questions (still unresolved). [E] If a note is ambiguous and you don't see where it fits, put it in a final "unclassified" block instead of inventing a category. Notes: [paste here] “

A case of iteration: a prompt that didn't work the first time

The v1 of prompt #3 (LinkedIn post) was this:

“ Write a LinkedIn post about a project I did. Make it sound authentic with a good hook. “

The output: a post with "here's a story," opening emoji, generic hook ("everyone talks about X, but…"), three predictable hashtags. In other words, exactly what I wanted to avoid.

The v2 — the one above — added three things: concrete data in [C] (client, industry, problem, result), a numeric constraint in [F] (180 words, 3 paragraphs), and a banned-phrases list in [E]. The change with the biggest impact: the banned list. I wasn't asking the AI to be creative; I was taking away the crutches it defaults to.

Practical rule I learned there: when a prompt gives generic results, don't add "be original." Name what you don't want.

To wrap up

Which of your repeating tasks are you going to turn into a prompt this week? Pick one — the one that makes you wrinkle your forehead on Mondays — and spend 20 minutes saving it properly. That investment pays itself back in two weeks.

If you want to revisit the CAFÉ base, go back to The CAFÉ Method. If you're still uncertain about what a prompt is, start with What is a prompt.

In a 2024 Q&A, Sam Altman said, with the casualness of someone who believes what he's saying is obvious, that the era of prompt engineering would die: models would get so good at guessing the intent behind a badly written request that "knowing prompting" would stop being a differential skill. It was, more or less, the same argument made years earlier about Google search: semantic search would make knowing how to write a good query irrelevant.

Three years later, models are indeed better at guessing. And at the same time — a paradox worth looking at closely — the people getting the most value out of these tools are still the ones with a library of their own prompts, built for their case, maintained over time. The AI got better at guessing. That doesn't replace knowing how to ask.

What Altman didn't foresee — or foresaw and underestimated — was that the problem wasn't writing prompts. The problem was managing reusable decisions inside AI-assisted workflows. And that problem doesn't get solved by better guessing; it gets solved with architecture.

The library as a layer, not a collection

A mature prompt library isn't a list. It's an abstraction layer over the model, with four components that get designed together.

Component 1: atomic prompts — single-task prompts, well-built with CAFÉ, with explicit variables. They're the base units. A good atomic prompt has clear input, predictable output, and documented behavior when something falls outside its comfort zone ("mark [ambiguous]", "return empty JSON", "ask for clarification").

Component 2: chained prompts — chains where the output of one prompt is the input of the next. The classic pattern: research → synthesis → drafting → review. Each link is an atomic prompt; the glue is a format contract between them (typically JSON or structured markdown, never free prose).

Component 3: meta-prompts — prompts that generate or improve other prompts. They're the reflective layer. A typical meta-prompt: "given this prompt that's producing generic results, identify which CAFÉ component is weakest and give me a v2 with that part reinforced." Meta-prompts don't scale your output; they scale your output of prompts.

Component 4: evaluation prompts — prompts designed to judge outputs from other prompts. This is the component that gets overlooked most. If your library has fifty prompts and you don't have a systematic way to know which ones perform well in which context, you don't have a library: you have an archive.

Workflows with prompt-as-input

The most useful pattern out of everything I've learned in two years doing this is chained prompts with a format contract. I'll describe it with a concrete case from my work.

When I analyze a client proposal for a process audit, the chain runs in four steps:

Step 1 (extraction). Atomic prompt that takes the raw document and returns JSON with fields: stakeholders, processes mentioned, metrics present, ambiguities detected. Output required as JSON with a fixed schema. If a field isn't in the document, it returns null — nothing gets invented.

Step 2 (analysis). Prompt that takes the JSON from step 1 and returns, for each process, a hypothesis of opportunity with confidence level (high/medium/low) and evidence supporting the hypothesis (direct quotes from the original document, not paraphrase).

Step 3 (prioritization). Prompt that takes the hypotheses from step 2 and ranks them by estimated impact × implementation ease, returning top 5.

Step 4 (drafting). Prompt that takes the top 5 from step 3 and writes the reply email to the client, in the tone my library has saved as "my voice in professional email."

Four atomic prompts, four clear format contracts, a workflow that used to take three hours and now takes forty minutes. But the important thing isn't the time. The important thing is that every time I find an error in the final output, I know exactly which step it came from and I can improve that step without touching the others.

That's what Altman missed: the edge isn't writing a good prompt. It's designing a system where prompts can be debugged separately.

Meta-prompts: automating improvement

The most useful meta-prompt I have saved is this one — you'll recognize it because it's model-agnostic:

`` [C] I have a prompt I've used [N] times that's giving me results I'm not happy with. I'll paste the prompt and three representative outputs. [A] Do three things: (1) identify which CAFÉ component (Context, Action, Format, Example) is weakest; (2) list the implicit decisions the prompt leaves to the model that would be better decided explicitly; (3) give me a v2 of the prompt that fixes those decisions. [F] Output: three sections ("Diagnosis", "Implicit decisions", "Prompt v2"). Prompt v2 must be directly usable, not a description of what would change. [E] If outputs are generic, the problem is almost always Example. If they're long when you wanted short, it's Format. If they drift off-topic, it's Context. If they don't do what you asked, it's Action. Current prompt: [...]. Three representative outputs: [...]. ``

What this meta-prompt does is turn prompt rewriting — a task anyone can do badly on intuition — into a process with explicit diagnosis. You're not asking "make it better." You're asking for structured diagnosis and a reasoned rewrite.

Prompts for evaluating outputs

Systematic evaluation is what separates a toy library from a working one. The base pattern: for every important prompt, keep a companion evaluator prompt that judges outputs against explicit criteria.

Example: for the LinkedIn post prompt that appears in Easy and Normal, I keep this evaluator:

`` [C] You evaluate LinkedIn posts against five objective criteria. [A] Given the post below, return a table with a 0-to-2 score per criterion and a single-line justification. [F] Criteria: (1) hook without cliché (0=guru cliché, 2=sounds like a real person); (2) one idea per paragraph (0=muddled, 2=clean); (3) closes with a real question (0=rhetorical or marketing CTA, 2=invites a genuine reply); (4) absence of LinkedIn jargon (0="plot twist", "unpopular opinion", etc., 2=none); (5) appropriate length (0=too short or too long for the content, 2=tight). [E] Don't moralize, don't suggest improvements. Score and justification on one line per criterion. Total: simple sum. Post: [...] ``

With this evaluator running over the last twenty posts produced by the main prompt, I have an empirical distribution of which criteria fail most often. If criterion 4 (no jargon) is systematically failing, I know exactly which part of the main prompt needs to be tightened — in this case, expanding the banned-phrases list.

This isn't academia. It's the difference between a library that grows in a directed way and one that grows by piling up files you can't tell still work.

Anti-patterns: what not to do

Three patterns that show up again and again in badly built prompts, and that meta-prompts and evaluators don't always catch automatically.

"Make it good." The most common and the most lethal. Telling the model "make it nice" or "make it professional" gives it no information — it gives it an aspiration without a criterion. "Good" for you might mean concise; for the model, by default, it tends to mean long and dressed up. Correct replacement: observable metrics (word count, mandatory element list, banned element list).

The paragraph-prompt. Twenty instructions hanging off one long sentence, comma-separated. The model processes it, but applies the instructions unevenly — the ones at the end weigh less than the ones at the front. Correct replacement: visual structure with bullets or delimiters (CAFÉ with its four separated blocks is exactly this).

Missing output format. The subtle one. You ask for analysis of a document and get an essay when you wanted bullets; you ask for bullets and get prose when you wanted JSON. The model picks a default format if you don't pick one. Correct replacement: the F component of CAFÉ always includes an explicit structure spec — even if it's just "return the answer as three paragraphs" instead of saying nothing.

Ethan Mollick wrote in 2024, in his newsletter One Useful Thing, a line that captures the point well: prompting isn't a magical ritual, it's applied communication — and like all communication, it works better when instructions are specific and expectations are negotiated explicitly.

Editorial thesis

The public conversation about AI in 2026 keeps fixating on model drama — who shipped what, which benchmark who broke, when the next jump arrives. That conversation leaves out what actually differentiates the professionals producing consistent results from the ones still exploring.

It isn't "knowing prompting" as a generic category. That, effectively, is getting commoditized — models are better at guessing, as Altman predicted. What doesn't get commoditized is having your own archive of reusable decisions, built for your context, maintained over time, evaluated against explicit criteria.

The competitive edge in 2026 isn't the tool. It's the accumulation. The professional who's been saving and refining prompts for three years has built an asset the person just starting out doesn't have, even though both use the same model with the same capabilities. And that asset isn't imitable by reading "best prompts" lists — because what makes it valuable isn't the individual prompts, it's the curation over your own work.

So, an editorial commitment: we're going to keep a public CAFÉ Library, with tested prompts tagged by use case, updated with what's working in real work. If you have a prompt that's solving something well for you, send it to the IA en Primera Plana newsroom — we review it, document it with its CAFÉ, and if it holds up, it goes into the catalog with credit to the author.

A living library. Built among readers. Maintained against reality, not against benchmarks.