Prompt Library for Offline AI

This is a working shelf of prompts for offline models, grouped by job. Each one is meant to be copied, edited, and saved in your own notes. A line under the prompt says when it is realistic: any small local model, or a larger one. “7B+” means a typical 7 to 8 billion parameter instruct model. “13B+” means you will get fewer dropped rules on a larger local model, not that a smaller one is forbidden.

The patterns behind these prompts are explained in How to Write Better Prompts for Local AI Models. More task types are in Best Prompt Templates for Local LLMs. Build a blank version in the Prompt Template Generator.

Email

Ask for something specific

Any small model, if you list the points.

Write a work email. To: [name] Ask: [one ask] Facts they need: - [fact] - [fact] Rules: - Under 110 words - No “I hope you are well” - End with the ask, not a vague offer to chat Output only the email.

Decline without padding

Any small model.

Write a short decline. What you are declining: [request] Reason you may mention: [one reason] Alternative you may offer: [one, or “none”] Rules: - Under 80 words - Do not apologize more than once - Do not invent a policy Output only the message.

Follow-up after silence

Any small model.

Write a follow-up to an email sent on [date]. Original ask: [ask] Rules: - Under 70 words - Refer to the earlier email in one sentence - Repeat the ask once - No guilt and no joke Output only the email.

Code

Function with an edge case

7B+ for a single function. Paste less than a page.

You are a careful [language] developer. Write one function that [does X]. Edge case to handle: [case] Rules: - Names should say what they hold - A 2-line comment at the top, no other comments - No example usage Return only the function.

Commit message from a diff summary

Any small model. Do not paste a huge diff; paste your own one-paragraph summary of the change.

Write a commit message. Change: [what changed] Why: [why] Rules: - Subject under 50 characters, imperative mood - Body: at most 2 sentences - Do not mention “this commit” Output only the message.

Review comments, not a rewrite

13B+ if the snippet is more than ~40 lines. Otherwise 7B is fine.

Review the code. Do not rewrite it. Return exactly 3 bullets: - Bug or “no bug found” - Naming or clarity issue, or “none” - Smallest next change Code: [paste]

Notes and meetings

Decisions and actions

Any small model.

From the notes, make two lists: Decisions and Actions. Rules: - Do not add items that are not in the notes - Actions start with a verb - Unknown owner: “Owner: not stated” Notes: [paste]

What changed since last time

Any small model. Paste both versions, labeled.

Compare OLD and NEW. Return: - Added - Removed - Unchanged but easy to miss One line each. If a list is empty, write “none.” OLD: [paste] NEW: [paste]

Extraction

Flat JSON

7B+. Smaller models often add a sentence before the JSON.

Extract JSON with keys: vendor, invoice_date, total. Rules: - null if missing - total is a number with no currency symbol - No markdown fence and no explanation Text: [paste]

Names and dates only

Any small model.

List every person and every date in the text. Format: People: - name Dates: - date — what it refers to, in 5 words, or “not stated” Do not infer dates that are not written. Text: [paste]

Rewrite and translation

Shorter, same facts

Any small model.

Rewrite the text in about half the length. Rules: - Keep every number, name, and date - Do not add a moral or a recommendation - No heading Text: [paste]

Translation with frozen terms

7B+ for a page. Any small model for a paragraph.

Translate from [source] to [target]. Keep these strings unchanged: [term], [term], [URL] Rules: - Neutral tone - No translator’s notes - If a word is unclear, keep the original in brackets Text: [paste]

Support and customers

Reply that does not promise a refund

Any small model. Put the policy in the prompt so the model does not invent one.

Reply to the customer. Policy you must follow: [one sentence, for example “We replace damaged items within 30 days. We do not refund shipping.”] Customer message: [paste] Rules: - Under 100 words - Say what you can do - Do not offer anything the policy does not allow - No fake empathy paragraph Output only the reply.

Turn a complaint into a ticket

Any small model.

Turn the complaint into a ticket with these labels: Issue: Product: When: Asked for: Rules: - One line per label - Quote the customer’s request, do not upgrade it Complaint: [paste]

Study

Questions from your notes only

7B+.

Write 5 questions that can be answered from the notes. After each question, write Answer: and one sentence. Then quote a short supporting phrase. If the notes do not support a question, do not ask it. Notes: [paste]

Glossary

Any small model.

From the text, build a glossary of terms a newcomer would not know. Format: term — definition in 20 words or fewer, using only the text. Skip everyday words. Maximum 8 terms. Text: [paste]

Decisions

Two options, one constraint

7B+.

Compare A and B for [decision]. The priority is [one constraint, such as cost]. A: [facts] B: [facts] Return: - Three lines, each starting with a criterion - Then “Pick:” and A or B, plus one sentence Do not add option C.

Pre-mortem in four lines

7B+. Do not ask for a long story.

Assume this plan failed after one month. Plan: [one paragraph] Return exactly 4 lines: Likely failure: Early sign: What we did not check: Smallest test this week: Do not give a pep talk.

How to file a prompt that earned its place

After a prompt works twice at temperature 0.2, save it with three notes: the model name, the temperature, and the kind of input that broke it. A prompt that works on a paragraph may fail on a ten-page paste. That is a context problem, not a reason to throw the prompt away. Check size with the Token Counter and split the source.

If a saved prompt starts adding intros again, do not rewrite it from scratch. Add one last line: “Start with the first required label.” Then retest once.

Related guides and tools