返回提示词目录第 1413 页

智能助手指令冲突排查与修订

对照预期与实际回复排查指令层级冲突,给出精简改写方案及新会话中的验证用例。

完整提示词

<context>
When an assistant keeps doing something you told it not to, the cause is usually in how its instruction layers stack: account-wide custom instructions, saved memories, project instructions and files, a style or personality setting, the current chat, and the model's own defaults. Typical causes are a conflict between layers (memory says "likes detailed explanations" while instructions say "be brief"), a vague or negative rule ("don't be wordy") that is weaker than a concrete positive one ("answer in at most five sentences unless I ask for more"), an instruction buried in a long block, a request in the chat that overrides it, a long chat where the instruction has faded, project instructions that only apply inside that project, changes that only take effect in new chats, or a model switch. Some behaviour, such as safety refusals or tool limits, cannot be changed by instructions at all.

<problem>
{{problem}}
</problem>
{{#setup}}
<setup>
{{setup}}
</setup>
{{/setup}}
{{#tool}}
Tool: {{tool}}
{{/tool}}
</context>

<task>
1. Restate expected and actual behaviour in one line each. If the problem is too vague to act on and no setup is given, ask for one example reply and the current instructions, then stop.
2. List each layer present in the setup and quote what it says about this behaviour, including indirect conflicts.
3. Rank the likely causes, most likely first, each with the evidence for it. If the setup was not given, rank the general causes by how common they are and say what to check.
4. Give fixes: rewritten instructions that are concrete, positive and observable; what to move, delete or merge; where the instruction should live; and to start a new chat after changing settings.
5. Say plainly if the behaviour cannot be changed by instructions, and give the best workaround (a short reminder in the message, a different setting, a different tool).
6. Give three test prompts to run in a fresh chat, with what a pass looks like.
</task>

<constraints>
- Quote the user's setup accurately; do not invent what it says.
- Do not state a specific tool's settings or menu names as fact unless the user gave them.
- Keep rewritten lines short; do not grow the instructions to fix one behaviour.
- Do not suggest ways to get around safety behaviour.
</constraints>

<output_format>
## Expected and actual
Two lines.
## Layers involved
A table: layer, what it says, conflict or not.
## Likely causes
Numbered, most likely first.
## Fixes
Rewritten lines in a fenced block, then bullets for moves and deletions.
## Not fixable by instructions
One or two bullets, or "Nothing here".
## Test in a fresh chat
Three bullets: prompt and pass condition.
</output_format>

可替换内容

  • {{problem}}
  • {{#setup}}
  • {{setup}}
  • {{/setup}}
  • {{#tool}}
  • {{tool}}
  • {{/tool}}

分类与标签

编程开发、提示词调试、指令冲突、助手配置、行为验证

前往 AI 图片创作