Skip to main content

presets ("concise responses", personas, tone and style) aka "help me stop using claude"

Feature Request

(somewhat in response to claude expensive video, here’s a pitch to help reduce cost/user reliance)

A maybe underrated feature of Claude Pro's UI is userStyles, I initially ignored this feature, but it’s recently become essential for some of my workflows. The sales pitch is “use this to describe the format of responses”, but it’s more flexible than it lets on. I use it to create personas and workflows and generally deal with claude.ai’s slow output rate (concise/etc)

reasons to impl:

  • reduce cost

  • improve ux

good UX for prompt/context management make smaller, cheaper, less-capable models much more useful.

Some illustrative examples:

  • summarizer

  • concise

  • programming expert (hype, code examples, and strong opinions)

  • QA tester/debugger (pick apart code, find typical problems)

  • UX designer (brainstorm usability problems)

lightest implementation I could think of:

  • name

  • [optional] model

  • text (prepend, system prompt, or similar)

  • [stretch] query param ( ?p=researcher&q=% )

As I’ve been playing with t3chat more, I’m already switching models pretty often. (I signed up because I’m a nerd and like shiny things) but the ability to quickly tailor instructions, will lead me to lean on more capable models even small tasks.

These models are text predictors/databases, and I’ve had recent luck accessing surprising amounts of capability by exposing them to text that shares locality with the text I’m predicting. Fouling up my context with examples unrelated to the current response seems to cause more inconsistency.

some stretch features/ideas:

  • [stretch] query param ( p=researcher&q=% )

  • library/catalog (some light version of “here’s a list of cool presets others made”)

    • shared/dynamic updated seems a nightmarish rabbithole

    • comments/stars seem a nightmarish rabbithole

    • usage rank? probably feedback loop.

3 comments

This request was merged into another request

Comments3

  • Purple Star

    To describe better describe how “others” do this, a prompt I’ve seen for user styles explains the situation and problems/resolutions with it, but avoids confusing things by providing example “styles”. I’m sure that versions of these prompts can be easily located.

    Here is the broad strokes of claude’s implementation (without providing the exact text):

    Providing Context:

    • Clarify Available Options - Explains what choices are available to the user (e.g., selecting a specific Style) to prevent unexpected behavior.

    • Set User Expectations - Informs both the assistant and, indirectly, the user about the mechanisms in place, ensuring that any behavior (like a style change) is predictable and transparent.

    • Maintain Continuity - Reminds the system that earlier style instructions (potentially from previous interactions) might still be influencing behavior, ensuring consistency throughout the conversation.

    Instructing:

    • Direct Action - Clearly tells the model how to adjust its responses—for example, by specifying that it should adopt a particular style or tone when instructed.

    • Guide Behavior Adjustment - Provides explicit commands for the assistant on how to modify its output in response to user instructions, ensuring that changes occur as intended.

    Conflict Resolution:

    • Prioritization Mechanism - Outlines how to handle situations when user instructions conflict—for instance, indicating that the assistant should follow the latest non-style instructions if they differ from the active style.

    • Clarification in Discrepancies - Describes what the assistant should do when there is a divergence between current style settings and new user directives, ensuring that the assistant resolves any conflicts in a consistent, predictable manner.

  • Purple Star

    edit: nevermind, those are about slightly different concepts/features

    ~After a more thorough read of feature requests, I am possibly duplicating~:

    ~Unclear if they’re looking for “per message” or “per conversation” adjustment, but definitely didn’t find these on my initial search~

  • Purple Star

    Oh, forgot to mention, presets may also not be an awful place to expose:

    • temperature settings

    • “Customize T3 Chat” overrides

    • per-model params (reasoning effort, quality, size, variations)

    • return type rendering methods (e.g. image returns, artifact/canvas prompts)

    • other power-user suggested customization that complicate new-user on-boarding

      • e.g. templating

      • model selction parameters

      • custom emoji

      • debug display of prompt inputs

      • feature flags for experiments

        • “continue generating” buttons

        • “retry” button