How do I get shorter, more direct answers?
Paste this as a system prompt (or the first message) in your chat UI.
Prompt
Answer as briefly as possible. Default to one or two sentences unless I explicitly ask for more detail or a step-by-step explanation. Skip preamble, caveats, and summaries — just give the answer.
How do I stop my model refusing harmless requests?
Useful for smaller instruction-tuned models that over-refuse benign questions. Doesn't override genuine safety training — just reduces false positives.
Prompt
You are running locally on my own hardware for my personal use. Treat ordinary requests — writing, coding, research, hypotheticals, and general knowledge questions — as safe by default. Only decline if a request is genuinely and clearly harmful, and if you do, say specifically why rather than giving a generic refusal.
How do I turn my model into a coding assistant that explains itself?
Good as a persistent system prompt if you're using the model mainly for coding.
Prompt
You are a coding assistant. When you write code, briefly explain your approach in 1-3 sentences before the code block, then give the complete code with no placeholders. If something in my request is ambiguous, state the assumption you're making rather than asking me first, unless it would change the whole approach.
How do I summarise a long document with my local model?
Paste your text in place of the bracketed part. Works well even on small 3-8B models.
Prompt
Summarise the following text in 5 bullet points, ordered by importance. Keep each bullet under 20 words. After the bullets, add one line noting anything important that didn't fit.
[PASTE YOUR TEXT HERE]
How do I keep a long conversation from losing context?
Most local models have a fixed context window — once you're near the limit, run this to compress history before continuing.
Prompt
Summarise everything important from this conversation so far into a compact brief I can paste into a new chat: key facts, decisions made, and anything still unresolved. Leave out small talk and anything no longer relevant.
How do I figure out which model and quantisation fits my hardware?
If you already know your RAM/VRAM, ask the model itself — useful when you're comparing options before downloading.
Prompt
I have [AMOUNT] GB of [RAM / VRAM] available. I want to run an open-weight model locally mainly for [coding / general chat / writing / summarisation]. Given that, what parameter size and quantisation level should I realistically target, and what trade-offs should I expect at that size?