You ask for the next step and get a page of explanation. A local model does not always know how much detail you need.
OGAD (Off Grid AI Desktop) gives you two ways to control that: ask for a specific answer shape, then set Max output if you also want a hard limit. If a reply is already too long, stop it and keep the text produced so far. You can do this with a downloaded local model, without sending the conversation to a cloud service.
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
These are desktop chat controls. Start with the core app and a local text model; the workflow does not need Pro work-history capture.
Ask for the result you want to read
A specific format gives the model a better target than “be concise.” Try a prompt like:
Give me the next three steps. Use one sentence per step. Start with the action. Leave out the introduction and conclusion.
Or, for a decision:
Recommend one option. Give the main reason and one limitation. Keep the answer under 120 words.
These are instructions, not guaranteed word counters. The model can still exceed them. But they help it spend its answer on useful content instead of using a token limit to cut off an otherwise long response.
Use a narrow question too. “What should I change first?” usually gives the model a clearer job than “Tell me everything about this.”
Add a hard response limit
In the chat, open Settings, then the model settings. Find Max output and choose a fixed value instead of Auto.
Auto lets the reply run until the model stops or fills the available context. A fixed Max output value places a cap on the response. In the current desktop settings, the selectable fixed values start at 2K tokens, subject to the context-window ceiling.
A token is a unit used by the model, not a word. A 2K-token cap does not mean 2,000 words, and it is still enough for a substantial answer. Use the prompt for a short paragraph or a small list; use the cap to prevent a much longer generation.
These settings are present in OGAD 0.0.51.
Keep three different limits separate
| Control | What it changes |
|---|---|
| Prompt instructions | The length and format you ask the model to follow. |
| Max output | The hard ceiling for the generated response. |
| Context window | How much material can fit in the model’s active context. |
Reducing the context window is not the right first step for a verbose answer. It can reduce how much of your input and conversation the model can use.
If you use a reasoning model with Thinking on, Thinking budget is a separate control for its reasoning tokens. The available budget is constrained by Max output. A very small total response budget can leave too little room for a useful final answer after reasoning.
For a simple request, start with a clear prompt and an ordinary local chat model. Adjust reasoning settings only when that workflow needs them.
Stop the reply without starting over
If the answer has already given you what you need, use the chat’s Stop control. OGAD keeps the partial text already produced in the conversation.
You can then ask for a tighter follow-up:
Turn the answer above into three short bullets. Keep only the actions I can take today.
Or, if the reply was cut off before the useful part:
Give only the conclusion that was missing. Do not repeat the earlier explanation.
Stopping leaves an incomplete answer if the model had more to say. Check that a code block, list, or instruction has not ended halfway through before using it. The follow-up is a new model response, so it can change wording rather than continue an exact hidden draft.
Try one question before changing every setting
- Load a local text model in OGAD.
- Ask a narrow question with a clear format, such as three one-sentence steps.
- If replies still run long, open the model settings and choose a fixed Max output.
- Ask again and check whether the answer is complete as well as short.
- Stop an unnecessary continuation and keep the useful text already on screen.
If the answer repeatedly ends mid-sentence, raise the cap and make the prompt more focused. A shorter useful answer is the goal; an arbitrary cutoff is only a limit.
Make the next answer easier to use
Download OGAD, load a local model, and give it one clear length instruction. Add a fixed response cap when you need it, and stop once you have the answer you came for.