How to control response length in HammerAI Desktop

By Arscelare
On this page

To get naturally shorter or longer replies in HammerAI Desktop, give the model a writing-length instruction and leave enough output allowance to finish it. The Max Response Tokens setting is a ceiling. Lowering it can cut a reply off before the model reaches a natural ending.

HammerAI can use local models or an external service through Proxy mode. The available controls and how they reach the model depend on your version and backend. The examples below are instructions to try, not tested presets or a promise of an exact word count.

A token ceiling does not set the prose length

HammerAI's Model Settings documentation calls the output control Response Token Limit (Max Response Tokens). Tokens are units of model text processing; they are not the same as words or paragraphs.

A maximum of 600 tokens does not ask for a 600-word reply. It permits generation up to the supported limit, while the model can stop earlier. Raising the ceiling allows a longer response but does not require one.

For naturally concise writing, tell the model what to include and where to stop. Use the output ceiling as a limit on generation, with enough room for the requested writing. A cutoff in the middle of dialogue is different from a finished short reply.

Put the instruction in a field the model receives

HammerAI's character documentation distinguishes the System Prompt, which guides model behavior, from the Creator's Note, which is informational text for people. It says the Creator's Note does not influence the model.

Put an ongoing writing rule in an instruction field that your setup sends, such as the character's System Prompt. Preserve the character's existing role instructions when adding it. Do not replace the whole prompt with a length rule.

Before changing several fields, save the original text. Check for an existing instruction asking for long descriptions, detailed internal monologue or a minimum length. Conflicting instructions make the test harder to interpret.

These field names come from public documentation. They do not establish the exact location of every control in your installed Desktop version or whether every local setting is forwarded in Proxy mode.

Try one clear short-reply instruction

For a dialogue-focused scene, this is an example instruction you can adapt:

Write one or two short paragraphs per reply. Include the character's immediate dialogue and one relevant action. Stop after that beat and leave room for my response.

Keep your current output allowance for the first test if it already permits complete replies. Judge whether the model ends naturally. Reducing the ceiling at the same time would hide whether the instruction changed the writing or the reply simply hit a limit.

If you prefer a word range, you can ask for one instead. Treat it as a target the model may miss, and inspect a few ordinary replies before deciding it works.

A short reply can still advance the scene. Specify the dialogue or action you want to preserve rather than asking only for fewer words.

Ask for fuller writing without asking for repetition

If the model stops cleanly but gives less detail than you want, a larger token allowance alone may have no effect. Give the additional space a purpose.

For example:

Write three to five paragraphs. Develop the character's reaction through dialogue, physical action and details they can observe. Advance the current exchange, then leave my character's response to me.

This is another example instruction, not an observed model response. Adjust the requested scope to your scene. A quiet exchange may need less space than a major confrontation.

Check the accepted messages and character examples too. HammerAI's character editor notes that a longer first message encourages longer replies. That is a style cue, not a guarantee. A short fresh test can help you distinguish the new instruction from a long-established pattern in the existing chat.

Diagnose a cutoff before changing more settings

What you seeWhat to check next
Replies finish cleanly but run too longThe sent writing rule, conflicting detail requests and the style in accepted history
Replies finish cleanly but feel too shortWhether the prompt asks for fuller writing and the output allowance leaves room for it
Replies repeatedly end mid-sentenceThe actual output limit and any reported finish reason; a connection interruption can also leave incomplete text
A tiny visible reply uses substantial outputWhether the selected model generates reasoning that the client hides
A rule has no apparent effectWhether its field reaches the model, including whether you used the display-only Creator's Note

A small diagnostic sequence is more useful than changing every sampler. Save the current setup, add one length instruction, and try a few replies. If they still cut off, check the backend's supported output allowance and request result before raising it.

Keep the model and other settings unchanged while comparing. Test generations through a metered service can incur charges even if you discard the replies.

Context size and reasoning can constrain the answer

Context Size controls how much input the setup can include. It is separate from the requested writing length. With a hosted model, the service's context and output limits still apply; changing a client number does not increase them.

Arscelare's model documentation, for example, specifies a 65,536-token context including the requested output allowance and an 8,192-token maximum output. It checks the fully rendered prompt and rejects overflow instead of trimming the chat or silently lowering the requested output.

Arscelare also defaults to medium reasoning. Reasoning and visible text share the output allowance, and both are billed at the output rate even if the client hides reasoning. A limit reached during reasoning can leave no visible answer. Tightening that limit to force a short reply can therefore produce the opposite of a useful concise response.

Those are Arscelare-specific rules, not universal HammerAI behavior. Check the documentation for whichever local model or hosted service you selected. Do not assume a client display toggle disables reasoning at the service.

If you want to use Arscelare, the HammerAI setup guide covers Proxy setup. Try the length instruction in a short scene first. Changing the model does not guarantee it will follow your preferred length.