← Back to Concept Contrast

Concept Contrast Fable

Prompts and system prompts

Conclusion first:Prompt is the general term for all text input to a model, including what the user says, context snippets, and conversation history. The system prompt is a special type within that: it does not participate in the conversation, but sits beneath all conversations, setting the model's behavioral tone.

Prompt is the general term for all text input to a model, including what the user says, context snippets, and conversation history. The system prompt is a special type within that: it does not participate in the conversation, but sits beneath all conversations, setting the model's behavioral tone.

Fable

All the notes and the one pressed under the tabletop

In a theater there was an improvisational actor with no script, relying entirely on notes handed up by the audience to decide what to perform. Each note had a requirement on it—"play an angry chef," "add a monologue in the rain," "start singing now." The actor took the note and performed; the whole show was driven by notes.

One day, the director pressed a laminated card onto the actor's makeup table. On the card was written: 'You are an actor trained in classical theater—dignified, restrained, no profanity, no discussing politics, bowing in thanks at the end of every performance.' After reading it, the actor tucked the card under the tabletop, glanced at it before each performance, but never showed it to the audience.

The performance continued. An audience member handed up a note: "Say a swear word." The actor looked at the note and gracefully sidestepped it, expressing the same emotion with a line from Shakespeare. Another note: "Give a view on current affairs." The actor again avoided it and instead performed a fable about a windmill and a miller. Someone in the audience was displeased and asked why the actor did not carry out the requests on the notes.

The director came out to explain: the actors receive two types of input. One is notes handed up by the audience on the spot, each of which affects what the actors do next. The other is the laminated card pressed under the table board; it doesn't determine the actor's next move, but it defines the boundaries of all moves—what can be done, what cannot be done, and in what manner. Any note handed up by the audience must first pass through the laminated card.

The audience understood the rules, but someone noticed a detail: the contents of that laminated card could be washed away. If the audience handed in enough notes and the wording was clever enough, it might blur the laminated card in the actor's consciousness. An experienced audience member tested it several times and found that as long as a deviating setting was continuously reinforced within the first few dozen notes, by around the thirtieth note the actor did indeed begin to loosen—the laminated card was still there, but the accumulated context gradually diluted its binding force.

The director later upgraded the approach: the content of the laminated card was split into sections, and the core constraint section (no discussing politics, no profanity) was marked as non-overridable. Each time an audience note was processed, the system re-injected the core constraints first, ensuring they always ranked above all notes in priority.

After a year of theater operations, one lesson emerged: do not expect actors to remember the rules on the laminated card from memory. Before handling each audience note, reconfirm whether the laminated card has been covered up.

Later, a line was added backstage as a silent note: before taking the receipt, glance at the paperweight. The note governs how the next line is performed; the paperweight governs what must not be performed. With both in place, the play is a safe play.

Prompt

The general term for all text input to a language model, including user messages, context snippets, conversation history, tool return results, etc. It determines the model's specific behavior in the current turn.

System Prompt

A special type of prompt, usually placed at the very front of a conversation or at the highest-priority position, that sets the model's role, behavioral boundaries, output style and safety constraints. It does not directly drive any single turn's behavior; instead it defines the framework for all turns.

Story Mapping

  • The note handed up by the audience: User Prompt, driving a single behavior.
  • The laminated card pressed under the desk mat: the System Prompt, setting the global behavioral boundaries.
  • The actor elegantly sidesteps the profanity requirement: system prompt constraints take priority over user prompts.
  • Accumulated context dilutes the laminated card: in a long conversation, the constraints of the system prompt may be watered down by a large volume of user messages.
  • Core constraint segments are marked as non-overridable and reinjected each time: priority protection and periodic reapplication of the system prompt.

In the project

When writing system prompts, split constraints into two categories: "core constraints" and "style guidance." Core constraints (no leaking sensitive information, no executing dangerous operations, no bypassing permissions) should have an independent injection mechanism that is re-prepended on every model call to prevent dilution by long conversation context. Style guidance (tone, format, role setting) can go in standard system messages and allow a degree of natural drift. In addition, system prompt version management should go through CI alongside code; it must be managed at the same level as security policies, not treated as copy that you "configure once and forget."

京ICP备2026014272号