These templates are written for local models such as Llama, Mistral, Phi, Qwen, and Gemma running in Ollama, LM Studio, Jan, or a similar runner. They use a short role, a single task, and hard constraints. That pattern is more reliable on a 7B or 13B model than a vague request like “make this better.”
Replace every bracket, then delete the brackets. Leave the rules in place. If the model still drifts, do these three things in order: move the rules above the source text, cut the source text so the instruction stays near the top, and lower temperature to about 0.2 in your runner. Count the prompt first if you are close to the context window — the Token Counter is an estimate, not a tokenizer, but it is enough to see when a prompt is getting long.
One task per prompt. “Write the email and also translate it and also make slides” is three prompts.
Name the output shape: bullets, a word limit, JSON keys, or “code only.”
Say what to leave out. “Do not add a preamble” stops a lot of local-model chatter.
Writing
Short article or post
Use this when you want a finished piece, not an outline. It works on 7B models if you keep the topic narrow.
You are a clear, concise writer.
Task: Write a blog post about [topic].
Requirements:
- Length: about 400 words
- Audience: [who]
- Tone: direct, no hype
- Structure: a 2-sentence opening, three sections with headings, a 2-sentence close
- Do not invent statistics
Output only the post.
Email that has to stay short
Local models pad emails unless you cap the length and list the points they are allowed to cover.
You are writing a work email.
To: [name or role]
Purpose: [ask / update / decline]
Points you may include:
- [point 1]
- [point 2]
Rules:
- Under 120 words
- No greeting longer than one line
- No “I hope this email finds you well”
Output only the email.
Product description
Write a product description for [product].
Facts you may use:
- [fact]
- [fact]
- [fact]
Rules:
- 70 to 90 words
- Second person (“you”)
- No unverified claims such as “best” or “guaranteed”
- End with one concrete use, not a slogan
Output only the description.
Coding
One function, one language
Ask for a function, not “an app.” Name the language version. Smaller models wander when the task includes a framework, tests, and a README at once.
You are a careful Python 3.11 developer.
Task: Write one function that [does X].
Input: [type]
Output: [type]
Rules:
- Include a docstring of 2 lines
- Handle this edge case: [case]
- Do not add a main block or example usage
- Return only the function
Explain a bug without rewriting the file
You are reviewing code. Do not rewrite the whole file.
Code:
[paste]
Return exactly 3 bullets:
1. The most likely bug
2. Why it happens
3. The smallest change that would fix it
If you are not sure, say “not sure” in that bullet. Do not invent a stack trace.
Convert a snippet, not a project
Convert the following [JavaScript] function to [Python].
Rules:
- Keep the same behavior
- Use clear names
- Add one comment only if a line is non-obvious
- Return only the converted function
Code:
[paste]
Summaries and notes
Fixed bullet summary
Summarize the text in exactly 5 bullet points.
Rules:
- Each bullet is one complete sentence
- Keep numbers and names that appear in the text
- Do not add background that is not in the text
- No title and no closing paragraph
Text:
[paste]
Meeting notes into decisions and actions
Turn these rough notes into two lists: Decisions and Actions.
Rules:
- Decisions: only things the notes say were agreed
- Actions: start each line with a verb and a name if one is given
- If an owner is missing, write “Owner: not stated”
- Do not add new tasks
Notes:
[paste]
TL;DR plus what to ignore
Read the text and return:
1. A TL;DR of 40 words or fewer
2. Three details that matter
3. Two details that look important but are not needed for the decision
Do not recommend a product.
Text:
[paste]
Translation and rewriting
Translation that keeps names
Translate from [English] to [Spanish].
Rules:
- Keep product names, URLs, and code unchanged
- Use a neutral tone
- Do not explain the translation
- If a sentence is ambiguous, translate the most literal reading and add “[ambiguous]” after it
Text:
[paste]
Rewrite for a reading level
Rewrite the text so a busy non-expert can follow it.
Rules:
- About the same length as the original
- Short sentences
- Keep every number
- Do not add examples that are not in the source
Text:
[paste]
Change tone without changing facts
Rewrite the paragraph in a calmer tone.
Rules:
- Do not add apologies that were not there
- Do not remove any deadline or amount
- Keep it under [80] words
Paragraph:
[paste]
Analysis and decisions
Risk review
You are a practical reviewer, not a cheerleader.
Subject: [feature, plan, or draft]
Constraints: [time, budget, or tools]
Return:
1. Three risks, each in one sentence
2. One way to shrink the scope
3. A yes or no, with a one-sentence reason
Do not invent metrics.
Compare two options
Compare Option A and Option B for [decision].
Option A: [facts]
Option B: [facts]
Return a table in plain text with columns: Criterion, Option A, Option B.
Then one sentence: which option you would pick if [constraint] is the priority.
Do not add a third option.
Structured extraction
JSON with nulls
Tell the model the keys and what to do when a field is missing. Otherwise local models invent a plausible date or price.
Extract JSON with exactly these keys: name, date, amount, currency.
Rules:
- Use null when the text does not state the value
- amount is a number, not a string
- Do not add keys
- No markdown fence
Text:
[paste]
Checklist from a policy
From the policy text, list only the actions a person must take.
Rules:
- One action per line, starting with a verb
- Skip background and definitions
- If the text is unclear, write “Unclear:” and quote the phrase
Policy:
[paste]
Study and practice
Quiz from your own notes
Using only the notes below, write 5 questions and answers.
Rules:
- Mix 3 short-answer questions and 2 true/false
- The answer must be supported by the notes
- After each answer, quote the supporting phrase in quotes
Notes:
[paste]
Explain, then stop
Explain [concept] to someone who knows [prerequisite] but not [concept].
Rules:
- 150 words maximum
- One concrete example
- No history section
- End after the example. Do not ask a follow-up question.
When a template still fails
If the model ignores a rule, repeat that one rule as the last line of the prompt. Local models often follow the nearest instruction. If it adds a friendly intro, add “Start directly with the answer.” If it refuses a normal writing task, the prompt is probably too long or the runner’s system prompt is fighting yours — shorten yours and check the app’s default system message.
For a pass/fail check before you rerun, paste the prompt into the Prompt Optimizer. It will not call a model. It only flags missing role, missing format, and prompts that are very short or very long.
