On this page
- The skeleton behind every template that holds up
- Template 1: Meeting summary with owners and deadlines
- Template 2: Cold email that doesn't sound like a template
- Template 3: Code review you can act on
- Template 4: Pulling structured data out of a long document
- Template 5: Rewrite in my voice, using my own samples
- A before-and-after you can test yourself
- Claude Code prompt templates are a different animal
- What to skip
- Keeping templates somewhere you will find them
- Next step
A good Claude prompt template is a fixed skeleton with a few [placeholders] you fill in each time: the role, the task, the input, the rules, and the output format. Below are five you can paste today, a before-and-after showing what the structure changes, and a plan for keeping templates from rotting in a notes app.
These are built for Claude, but they transfer to ChatGPT and Gemini with small edits. If you want a broader set of starters, we also have 7 Claude templates you can paste today. This post goes the other way: it explains why the skeleton works so you can write your own.
The skeleton behind every template that holds up
Every template below follows the same five parts. Once you see them, you can rebuild a template for any task in a couple of minutes.
Context. Who you are, who the output is for, and why it exists. One or two sentences.
Task. One verb-led instruction. Summarise, rewrite, review, classify. Not three at once.
Input. The material Claude should work on, fenced off so it can't be confused with your instructions.
Constraints. What to avoid, what to keep, what to do when information is missing.
Output format. Length, structure, tone. Be explicit, because Claude will otherwise pick its own.
The fencing in step 3 is the part people skip. Claude handles XML-style tags well, so wrapping pasted material in tags like <transcript> or <draft> is a cheap habit that prevents a lot of confusion. Pasted emails often contain instructions of their own, such as a line reading please reply by Friday, and tags help Claude treat that as content, not as a command.
The constraints in step 4 should include an exit. Tell Claude what to do when the input doesn't contain what you asked for. Without that line, it tends to fill gaps with something plausible.
Template 1: Meeting summary with owners and deadlines
Use this for call transcripts and rough notes. The key move is the rule against inventing owners. Meeting notes rarely say who agreed to what, and a summary that guesses is worse than none.
You are helping me write up a meeting for people who were not there. Summarise the transcript below. <transcript> [paste transcript or notes] </transcript> Rules: - Only include decisions and actions that are stated in the transcript. - For each action, give the owner and deadline if they are stated. If not, write 'owner: not stated' or 'deadline: not stated'. Do not guess. - Ignore small talk and repeated points. Output format: 1. Decisions (bullets, one line each) 2. Actions (bullets: action, owner, deadline) 3. Open questions (bullets) Keep the whole summary under [number] words.
The 'not stated' phrasing matters. It gives Claude an approved way to be incomplete, so it doesn't feel pushed to complete the pattern.
Template 2: Cold email that doesn't sound like a template
Most cold email prompts produce the same opener: I hope this email finds you well. This one blocks that by supplying a real detail about the recipient and banning the usual openers.
Write a cold email from me to a stranger. About me: [your role, company, one line on what you do] Recipient: [name, role, company] Something specific I noticed about them: [a post, a product launch, a talk, a hiring page] What I want: [one clear ask, e.g. a 15-minute call] Why they should care: [one concrete benefit, no more] Constraints: - Under [100] words. - Open with the specific thing I noticed, not with an introduction of myself. - No phrases like 'I hope this finds you well', 'quick question' or 'touch base'. - One ask only. Plain language. No exclamation marks. - Write the subject line too, under 6 words. Give me two versions with different tones: one direct, one warmer.
Asking for two versions is deliberate. You'll usually pick one and borrow a sentence from the other, which is faster than iterating on a single draft.
Template 3: Code review you can act on
Unstructured review requests get you a polite essay. Asking Claude to rank findings by severity and quote the lines involved gets you a list you can work through.
Review the code below as a senior [language] engineer. Context: [what this code does, where it runs, and anything you already know is a trade-off] <code> [paste code] </code> Check for, in this order: 1. Bugs and incorrect behaviour 2. Security problems 3. Error handling gaps 4. Readability issues that would slow down the next person For each finding give: severity (high / medium / low), the exact line or snippet, why it is a problem, and a suggested fix. Do not comment on formatting or naming style unless it causes a real bug. If you find nothing in a category, say 'none found' rather than inventing something.
Template 4: Pulling structured data out of a long document
This one is for contracts, reports, and long email threads. Because the output is meant to be parsed or pasted into a spreadsheet, the format section does the heavy lifting.
Extract the information below from the document. <document> [paste document] </document> Fields to extract: - [field 1] - [field 2] - [field 3] Rules: - Quote the source sentence for every value you extract. - If a field is not in the document, return null for that field. Do not infer it. - If a field appears more than once with different values, list all of them and flag the conflict. Output as JSON with the field names above as keys, and a 'source_quote' entry next to each value.
The quote requirement is a cheap audit trail. You can check any value in seconds, and Claude is less likely to drift from the text when it has to point at it.
Template 5: Rewrite in my voice, using my own samples
Telling Claude to be casual but professional gives you generic output. Giving it two examples of your own writing gives you something close. This is few-shot prompting in practice, and we cover the technique in more detail in our few-shot prompting guide.
Rewrite the draft below so it sounds like me. Here are two samples of how I write: <sample_1> [paste a paragraph you wrote] </sample_1> <sample_2> [paste another paragraph you wrote] </sample_2> <draft> [paste draft] </draft> Match my sentence length, vocabulary and level of formality. Keep every fact and number from the draft exactly as written. Do not add claims. Return only the rewritten text, then a two-line list of what you changed.
A before-and-after you can test yourself
Take a task most people do weekly: turning a messy Slack thread into a status update for a manager.
The usual prompt is this: summarise this thread for my manager. Claude will give you a friendly paragraph. It will probably include side comments, soften a blocker into a concern, and run to a length nobody asked for.
Here's the same task with the skeleton applied:
I'm a project lead writing a weekly update for my manager, who reads it in under a minute.
Turn the Slack thread below into a status update.
<thread>
[paste thread]
</thread>
Rules:
- Lead with anything blocked or at risk. State blockers plainly, don't soften them.
- Only include what is in the thread. If the thread doesn't say when something ships, write 'date not confirmed'.
- Skip side conversations.
Format: three headings (Blocked, In progress, Done), at most three bullets each, one line per bullet.What changes in the answer: the blocker moves to the top, the vague 'we're hoping to ship soon' becomes 'date not confirmed', and the whole thing fits on one screen. You haven't made Claude smarter. You've removed the decisions it was making for you.
Claude Code prompt templates are a different animal
Templates for Claude Code, the terminal-based coding tool, work best when they split into two layers. Standing instructions about your project, such as the test command, the folder structure, and the style rules, belong in a project instructions file that Claude Code reads at the start of a session. One-off task prompts are the templates you reuse.
A task template for a bug fix looks like this:
Fix this bug: [one-sentence description of the wrong behaviour] Expected: [what should happen] Actual: [what happens, including any error message] Where I think it is: [file or function, or 'unknown'] Before changing anything, find the cause and explain it in two or three sentences. Then make the smallest change that fixes it. Run [test command] and report the result. Do not refactor unrelated code.
The explain-first step catches wrong diagnoses before they become wrong edits. Suppose you want to understand where instructions should live in agent setups, AGENTS.md, Skills, or MCP cover the options, and how that works varies by tool and version, so check the current docs for your setup.
What to skip
Not every template needs all five parts. A one-line translation or a quick rename doesn't need a skeleton, and wrapping it in one wastes your time. Use a template when you'll run the task again, when the output feeds into something else, or when a colleague will need to get the same result.
Also skip long lists of rules. Five clear constraints beat fifteen overlapping ones, and contradictory rules produce erratic output. If a template keeps failing in the same way, add one targeted line about that failure rather than rewriting the whole thing.
Keeping templates somewhere you will find them
The real cost of templates isn't writing them. It's that the good version lives in a chat from three weeks ago and the version your colleague uses is an older copy. A few habits fix most of it:
One home. Put every template in a single place with a name that says what it does, like Meeting summary - owners and deadlines, not Prompt 7.
Note the model. A template tuned on Claude may behave differently on ChatGPT or Gemini. A one-line note saves someone a confused afternoon.
Change one thing at a time. If you edit three parts of a template at once, you won't know which edit helped.
Keep the old version. When an edit makes things worse, you want to go back. Prompt version control exists for exactly this, and in PromptPal you save a template once and keep its history instead of juggling copies.
A practical workflow: draft the template in a chat, run it on three real inputs, fix whatever failed, then save it with the placeholders intact. Three real inputs is the minimum. A template that works on the one example you wrote it for hasn't been tested.
Next step
If you'd rather start from something that already works than build from the skeleton, browse the public prompt library, copy a template for your role, and save your edited version to your own library.
Stay Updated
Get the latest blog posts and AI prompt engineering tips delivered to your inbox.




