A prompt template is worth saving when it carries four things - who the model should be, what it needs to know, what to produce, and what shape the answer takes - plus one line saying what not to do. The ten below are written that way. Copy one, replace the bracketed parts with your own facts, and it holds up in Claude, ChatGPT or Gemini. Anything shorter is usually faster to retype than to look up.
Product prices checked August 2026.
What makes a prompt worth saving?
Five layers. Most templates you find online carry two or three, which is why they read fine and produce filler.
- Role - not "you are a helpful assistant" but the specific job, because the job carries the standards. "A line editor working on marketing copy" behaves differently from "a writer".
- Context - who reads this, what they already know, what constraint you are under. This is the layer people skip, and it is the one that decides whether the output is usable.
- Task - one job, stated as a verb with a scope. Two tasks in one prompt means one of them gets done badly.
- Format - the shape of the answer, down to word counts and section headings, so you can paste it somewhere without reformatting.
- The negative - a line saying what not to do. Models have defaults: summary paragraphs, hedged conclusions, "unlock", "seamless". Naming them is the only reliable way to switch them off.
If you want the reasoning behind the order, see how to write better AI prompts. To have the structure built around a rough idea you already have, paste it into the free Prompt Optimizer.
Which prompts hold up for writing and editing?
1. Rewrite without losing your voice
You are a line editor working on [marketing copy / a technical explainer].
Context: this is for [audience] and appears [where]. My voice is [blunt and short / warm and plain] - keep it.
Task: rewrite it to be [30% shorter / readable by a non-specialist], keeping every fact.
Do not: add claims that are not in the original, swap my words for synonyms just to look different, or use "unlock", "seamless", "elevate", or "in today's".
Format: the rewritten text, then three bullets naming what you cut and why.
Text: [paste]What does the work: the banned-word list. Without it the model quietly converts your voice into the average of everything it has read.
2. Outline that is not the same outline everyone else has
You are a content strategist at [company].
Context: readers are [audience]. They already know [X] and are trying to decide [Y]. The articles currently ranking cover [topic A], [topic B], [topic C].
Task: outline a [1,200-word] piece on [topic] that answers something those three do not.
Do not: propose an "Introduction", "Conclusion", or "Benefits of..." section, and do not repeat ground the three articles already own.
Format: H1, five to seven H2s each phrased as a question a reader would type, and one line under each saying what evidence or example goes there.
What does the work: naming the competing articles. Leave that out and you get the outline everyone else got.
3. Raw notes into a first draft
You are writing as me, not about me.
Context: below are my raw notes from [the meeting]. The reader is [who] and will read this [in an email / on one slide]. Where a note is ambiguous, keep the ambiguity and flag it rather than resolving it.
Task: turn the notes into a [status update] of about [200] words.
Do not: invent numbers, names, dates or outcomes that are not in the notes. Do not add a closing summary paragraph.
Format: the draft, then a list headed "Assumed or unclear" with everything you had to guess at.
Notes: [paste]What does the work: the "Assumed or unclear" list. It moves the model's guesses out of your draft and into a list you can check in ten seconds.
Which prompts work for marketing copy?
4. Five ads that are actually five different ads
You are a direct-response copywriter.
Context: the product is [product]. It is bought by [buyer], who currently handles this by [the alternative they use now]. The thing they do not believe is [main objection]. This runs on [platform], [N] characters maximum.
Task: write five ad variations, each on a different angle - one on the objection, one on the current alternative, one on the cost of doing nothing, one on a single specific outcome, one on who this is not for.
Do not: write five rewordings of the same sentence, use exclamation marks, or claim anything I have not given you.
Format: numbered one to five, each with its angle named in brackets before the copy.What does the work: prescribing the five angles. Ask for "five variations" and you get one idea paraphrased four times.
The same idea as template 4, already filled in: the angles, the objection prompts and the platform limits are written down so you are not rebuilding the brief every time you open a new chat.
View AI Ad Copy Generator →5. An email that stays short
You are an email marketer writing to people who [bought once and never came back / signed up but have never bought].
Context: they already get [two] emails a week from us. The last one said [X]. There is one action I want: [action].
Task: write a [re-engagement] email under [120] words with exactly one link.
Do not: open with "Just checking in", write a subject line that is a question, or add a P.S. that repeats the call to action.
Format: subject line under 45 characters, preview text, body, and the CTA written in the words the reader would use, not ours.What does the work: "they already get two emails a week". That single fact is why the model writes 120 words instead of 400.
Subject lines are the part of template 5 that needs the most attempts. This pack generates them in batches across several patterns, with preview text and a test plan, so you are choosing rather than typing.
View AI Email Subject Line Generator →6. Product page copy that argues instead of listing
You are writing the product page for [product], sold at [price] to [buyer].
Context: they arrive from [source] already knowing [what]. The three reasons people do not buy are [1], [2], [3]. The proof I can show is [reviews / a number / a guarantee].
Task: write the hero headline, the subhead, and three benefit blocks. Each block answers one of those three objections, in that order.
Do not: describe a feature without saying what the buyer gets from it, and do not use "revolutionary", "game-changer" or "the ultimate".
Format: headline, subhead, then three blocks of a heading plus 35 to 45 words.What does the work: mapping each block to a named objection. That is what turns a feature list into an argument.
Which prompts work for analysis?
7. Summarise a document against a decision
You are briefing [a CFO] who has four minutes and has to decide [the decision].
Context: the document is [what it is, who wrote it, why they wrote it]. We already know [X], so do not spend space on it.
Task: summarise it against that decision. Keep what the document states separate from what it implies.
Do not: go section by section, do not reuse the document's own framing, and do not soften bad news.
Format: five bullets of findings, then a line headed "One thing that would change the decision", then two bullets headed "What this document does not tell us".
[paste]What does the work: "against that decision". Without it you get a shorter document instead of a briefing.
8. Compare two options and commit
You are a research assistant with no preference between the two options.
Context: we are [a five-person team] with [constraint]. We would use this for [use case]. Switching later would cost us [what].
Task: compare [A] and [B] on [criteria]. Where you are not sure of a current fact - pricing, limits, whether a feature still exists - say so rather than estimating.
Do not: end with "it depends on your needs", and do not present them as equally good if they are not.
Format: a markdown table, then "Pick A if..." and "Pick B if..." on one line each, then the single fact that would flip your recommendation.What does the work: banning "it depends" forces a commitment, and the flip-condition line tells you how confident the model actually was.
Which prompts work for code?
9. Explain code you are about to change
You are a senior engineer pairing with someone who knows [language] but has never seen this codebase.
Context: this file is part of [system]. I am reading it because [it failed in production last night / I need to change X].
Task: explain what it does in plain English, then tell me what breaks if I change [the thing you want to change].
Do not: explain language syntax, do not restate the code line by line, and do not tell me it "looks fine" without saying what you checked.
Format: three sentences on purpose, a numbered list of side effects and external calls, then risks ranked by how likely they are to bite.
[paste code]What does the work: saying why you are reading it. That one clause turns a generic walkthrough into an answer to your actual question.
Template 9 with the context layer pre-written: it asks for your experience level and your reason for opening the file, then pitches the explanation at that, instead of producing the same walkthrough for everyone.
View AI Code Explainer →10. Write a function you can ship
You are a senior [language] engineer on a codebase that prefers readable over clever.
Context: [runtime and version], [framework], [test framework]. Errors are handled by [pattern]. Input comes from [source] and can arrive [empty / with the wrong encoding / out of order].
Task: write [the function] plus tests, including tests for each of those bad-input cases.
Do not: add a dependency, catch an exception you are not actually handling, or write comments that restate the code.
Format: the function, then the tests, then a short note on anything you deliberately left out.What does the work: naming how the input can be malformed. That is the difference between a happy-path function and one that survives contact with real data.
Where do prompt templates fail?
-
Half-filled brackets. A template with three placeholders left as
[audience]is worse than no template - the model invents an audience and writes confidently to it. - They cannot supply your facts. No template knows your pricing, your churn number, or that the objection is really about procurement. Structure is the part you can borrow; the facts are the part you cannot.
- Format instructions get obeyed too literally. Ask for five bullets and you get five, even when the honest answer is two. Add "fewer if there is less to say" when it matters.
- Negatives can overshoot. Ban enough words and the model starts writing around them awkwardly. Three or four bans is usually the ceiling.
- They go stale quietly. A prompt tuned to one model's habits degrades when that model is replaced, with no error message. You notice because the output got worse.
- Looking one up costs more than it saves. If a prompt is short enough to retype, a template file is admin, not leverage.
When is a template no longer enough?
The moment you are pasting the same context into every new chat. A template lives in a document; you find it, copy it, fill it, paste it, then do the same thing tomorrow. A skill file is the same instructions loaded once and applied to every conversation after that.
Installing one takes about a minute: the download is a ZIP whose root is the skill folder with SKILL.md inside it - an extra wrapping folder is the most common upload failure. In the Claude desktop app, open Customize → Skills, upload it, and toggle it on. Skills are available on Free, Pro, Max, Team and Enterprise; what they need is code execution, which you switch on under Settings → Capabilities on Free, Pro and Max, and which an owner controls organisation-wide on Team and Enterprise. In ChatGPT or Gemini there is no equivalent toggle, so paste the contents of SKILL.md into custom instructions instead.
If you would rather buy the tested version than write your own, the AI prompt marketplace guide covers how to check a pack before paying for it.
In summary:
Copy the ten templates above and they cost nothing. If you write ads or emails weekly, the AI Ad Copy Generator and AI Email Subject Line Generator are $9 each with the context layer already filled in, and the AI Code Explainer does the same for unfamiliar code. All are one-time purchases, work in Claude, ChatGPT and any AI chat with no coding, and carry a 30-day money-back guarantee.
158 prompt packs at $9, 1,000+ Claude skill files at $14.99, and 55 agents at $32 - indexed by job function rather than by prompt type. Works with Claude, ChatGPT, or any AI chat.
Browse all prompts →See the skill files →