Preset Author Guide

This page is for authors who are already comfortable with presets, PromptManager, and prompt order.

Agent Mode does not move preset authors away from the existing creation system. It keeps one of the most valuable parts of that system: you can still decide how context is arranged, where character information sits, how World Info before and after relate to the rest of the prompt, and which instructions stay close to chat history.

The difference is that Agent adds two dedicated PromptManager components:

Agent System Prompt
Agent Results

Both can be moved in the preset order. Understanding what they mean is the key to making presets and Agent Profiles work together steadily.

Three Layers

Presets, Profiles, and SKILLS have different jobs:

Layer	Main responsibility
Preset	Controls the order of ordinary prompts and Agent-specific position blocks
Agent Profile	Controls the Agent system prompt, tools, SKILLS, workspace permissions, presentation, and output policy
SKILLS	Hold longer writing methods, examples, format rules, and reference material

A gentle rule of thumb: the preset decides where things go, the Profile decides what the Agent may do and how it should work, and SKILLS hold material the Agent can read when needed.

Agent System Prompt

Agent System Prompt is a position anchor in the preset. It is not the content source.

When Agent Mode starts, the frontend places an internal marker at this position. The backend runtime then replaces that marker with the system prompt resolved from the current Agent Profile. In other words, the real Agent system prompt comes from instructions.agentSystemPrompt in the Profile, not from text written into this preset block.

This has a few practical consequences:

• Do not write behavioral instructions into the preset's Agent System Prompt block.
• To change Agent behavior, edit or distribute an Agent Profile.
• This position must exist. If it is missing, the Agent run fails clearly instead of falling back silently.
• Older presets that do not have this position are given a default reference when possible.

The default order places it before Main Prompt. That is usually a safe choice, because the Agent work brief enters the context before the ordinary main prompt.

If you have a clear reason to move it, move it by intent:

Placement	Good fit
Early, near Main Prompt	The Agent work brief should frame the whole writing task first
After character, World Info, or persona material	The Agent should see core setting first, then receive this Profile's work brief
Near chat history	The Agent action rules should sit close to this turn's conversation context

Avoid moving it often. Its position changes how the Profile system prompt relates to the rest of the context. It is best to change one factor at a time and test with the same chat.

Agent Results

Agent Results brings previously committed Agent results back into a later Agent prompt.

It does not place raw tool logs, drafts, or hidden workspace files into the context. It reads Agent commit metadata and message text saved on chat messages, then places them as system messages at the Agent Results position.

It helps later Agent runs know:

• Which chat messages came from Agent.
• The earlier runId, workspaceId, checkpointId, and commit information.
• The final text that was committed to chat.

Agent Results can be moved. It can also be disabled when you intentionally do not want later Agent runs to see previous Agent results.

Common placements:

Placement	Good fit
Near chat history	Previous Agent output should behave like part of the conversation context
Before Post-History Instructions	Final post-history rules should still constrain how Agent uses these results
Disabled	The play style avoids previous Agent results, or old results tend to cause repetition

Agent Results uses context budget. Long or older Agent results may be omitted when the budget is exhausted. This is normal budget behavior, so it should not be treated as unlimited memory.

Short durable facts belong in persist/ or a future memory mechanism. Longer writing rules belong in SKILLS.

Relationship to Agent Profiles

The relationship between a preset and a Profile can be read this way:

The preset decides where the Profile system prompt enters the prompt.
The Profile decides what that system prompt says and what the Agent may do.

instructions.agentSystemPrompt in a Profile commonly has two states:

Value	Meaning
null	Use TauriTavern's runtime default Agent system prompt
Non-empty string	Fully replace the default system prompt

If you want the default behavior, keep it as null. If you provide a custom prompt, write it as a complete work brief that can stand on its own, not as a partial patch.

At the current stage, Profiles do not switch model or provider for the user. model.mode still follows the current prompt snapshot, and preset references are mainly for record and validation rather than rewriting context. Preset authors should still remind users to choose suitable API, model, and context settings.

Distribution Advice

If a preset is adapted for Agent, consider distributing:

• The preset itself.
• The recommended Agent Profile.
• SKILLS paired with that Profile.
• A short note explaining how to use it with Agent Mode off, and which Profile to select when Agent Mode is on.

Do not assume the user already has a Profile or Skill with the same name installed. It is steadier to list dependencies clearly and keep the ordinary preset usable when Agent Mode is off.

Before Publishing

A small check before publishing helps:

Check	Expected result
Agent Mode off	Agent-only blocks do not affect ordinary generation
Agent System Prompt	The position exists, has not been removed, and sits where you intend
Agent Results	The position matches your intent. If disabled, that is deliberate
Profile	instructions.agentSystemPrompt is written in the Profile, not in the preset marker
SKILLS	Long material is not piled into the system prompt, and Agent can read it as needed
Tool policy	Allowed tools match what the system prompt asks Agent to do

Preset authors do not need to turn Agent into a second complicated system. The best adaptation is often just two clear positions, one well-bounded Profile, and a few readable Skill files.