Write better prompts for small local AI models

Ask for one meaningful transformation

Rewrite this email in a calmer tone is easier to follow than act as a strategist, editor, lawyer, and marketer at once. Choose the immediate job and describe the input. If the work has several stages, complete them in separate turns with a short review between stages.

For example, first ask the model to identify unclear sentences, then choose which ones need rewriting. This preserves your control over meaning and reduces the chance that a broad request changes facts, tone, and structure all at once.

Put important constraints next to the task

If dates and prices must remain unchanged, say so explicitly and keep those facts easy to see. Separate source text from instructions with labels. Avoid a long list of personality instructions that competes with the actual work.

A small example can clarify the desired format, but make sure it represents the task. An example with different requirements can confuse the model. State whether it is a formatting example or factual source material.

Example prompt

Task: Rewrite the message below to be concise and polite.
Preserve: Every date, amount, name, and commitment.
Do not: Add promises or new facts.
Output: One revised message under 120 words, followed by any ambiguity you noticed.
Source message: [text]

Give feedback about a specific failure

If the answer is too long, identify the target length. If it changed a fact, point to the original and the incorrect version. Asking the model to try harder or be smarter does not tell it what to correct.

Keep revisions focused. After several rounds, start a fresh chat with the latest accepted text and the remaining task. A long conversation containing discarded instructions can make the current request harder to follow. See the context guide for how history affects a reply.

Know when prompting is not the bottleneck

A clearer prompt cannot give a model tools it does not have, current information it has not received, or unlimited reasoning ability. If the task needs web research, reliable calculation, or execution, use those tools separately and provide the relevant evidence.

Try a larger model only if your device can run it comfortably, and compare results on the same small task. If an answer remains unreliable, simplify the task or use a more suitable method. Effective prompting is about making success checkable, not finding a magic phrase that guarantees correctness.

About this guide

Original examples are illustrative. This is AI-assisted content maintained for GPT0, not an independent product benchmark. Read our editorial approach.