Write prompts that work on private AI

A prompt is the message you type to tell the AI what you want. Private AI models are smaller than the big online tools, so a prompt that works fine in ChatGPT can get a muddled answer on your own computer. This guide gives you a simple five-part pattern that gets clear, repeatable results from private AI, and shows you how to save your best prompts so you never write them from scratch again.

Before you start

• A private AI app with a model downloaded, such as LM Studio, Ollama or AnythingLLM.
• One task you do often, like summarising notes, replying to a type of email or rewriting product descriptions.
• A real example of that task’s input, such as a set of notes or an email.
• About 20 minutes.

Want to do it yourself?

Difficulty

4 out of 10

Estimated completion time

20–40 minutes

Skills you’ll need: Describe one job, separate instructions from source material, and test the prompt on several examples.

I’ll do it myselfI’d rather get help

How this helps your brand

A saved prompt gives recurring private writing and summary tasks a consistent structure. Testing missing and unusual inputs helps you spot gaps before a draft becomes a customer-facing message.

1. 1

   Step 1: Name one job

   Start with a single, clear instruction. “Summarise these meeting notes” works. “Summarise these notes, then draft an email, then suggest a social post” often gets a confused answer from a local model.

   If you need several things, ask for them one after another in the same chat.

2. 2

   Step 2: Say who it’s for and what it should look like

   Tell it who will read the result, and exactly what shape you want back. The more specific, the better.

   • Reader: “for a customer who isn’t technical” or “for my team”.
   • Format: “five bullet points”, “a short email with a subject line”, “a table with two columns”.
   • Length: “under 100 words” or “no more than three sentences”.
3. 3

   Step 3: Set the rules

   Local models are more likely to fill gaps with guesses. Tell it what to do when information is missing, and what it must not do.

   • “Use only the information below.”
   • “If something is missing, write [missing] instead of guessing.”
   • “Don’t add prices, dates or promises.”
4. 4

   Step 4: Put your material last, clearly marked

   Put the instructions first and the notes or email at the end, under a clear label like “Notes:” or “Email:”. That way the model doesn’t confuse your instructions with the material it’s working on.

   Paste only what it needs. Long inputs slow local models down and make them more likely to lose track.

   Tip: If you want a particular style, add one short example of a good answer. One example usually helps more than a paragraph of description.

   Illustrated exampleBuild one prompt with five clear parts

   Example promptTeam meeting summary

   1 Task

   Summarise these meeting notes.

   2 For

   My team.

   3 Format

   Three bullet points, under 100 words.

   4 Rules

   Use only the information below. If something is missing, write [missing] instead of guessing. Don’t add prices, dates or promises.

   5 Notes

   Maya will send the revised quote. The team chose the blue packaging. The delivery date is still unconfirmed.

   Instructions above. Clearly labelled source material last.

5. 5

   Step 5: Test it on three examples, then save it

   Try your prompt on a normal example, one with something missing, and one that’s a bit unusual. If an answer goes wrong, change one line of the prompt at a time and try again, so you know which change helped.

   When it works on all three, save it. Most apps let you save standing instructions as a system prompt or preset. Otherwise, keep a note called “AI prompts” with each one under a clear name.

   Illustrated exampleCheck the same prompt against three kinds of input

   Same prompt. Three different sets of notes.

   1. Normal
      Notes in

      Maya will send the quote on Friday.

      Expected detail in summary

      Maya will send the quote on Friday.

      Keeps the stated person and date.

   2. Missing detail
      Notes in

      Maya will send the quote. No date agreed.

      Expected detail in summary

      Maya will send the quote. Date: [missing].

      Marks the gap without inventing a date.

   3. Unusual
      Notes in

      Friday was suggested, then rejected. No date agreed.

      Expected detail in summary

      Quote date: [missing]. Friday was rejected.

      Doesn’t turn a rejected suggestion into a promise.

   Save when all three workPreset or note name: “Team meeting summary”

   A test fails? Change one prompt line, then run the same three examples again.

Task: [one job, for example: summarise these notes]
For: [who will read it, for example: my team]
Format: [for example: five bullet points, under 100 words]
Rules: Use only the information below. If something is missing, write [missing] instead of guessing. Don’t add prices, dates or promises.

[Label, for example: Notes]:
[paste the material]