When to Reveal a Character's Secret: Writing Three-Layer Disclosure Conditions for AI Dialogue
You can split the same secret into three layers of information, each tied to observable evidence and relationship behavior: only when the player has seen something and done something does the character have a reason to speak more deeply. Chat turn count can help pace the rhythm, but it cannot by itself explain why a character is willing to disclose. Below, a fictional teaching example illustrates the approach. Lin Wan, a secondhand bookstore clerk, conceals the true sender of an old book; we split this secret into three layers.
Introduction
You can split the same secret into three layers of information, each tied to observable evidence and relationship behavior: only when the player has seen something and done something does the character have a reason to speak more deeply. Chat turn count can help pace the rhythm, but it cannot by itself explain why a character is willing to disclose. Below, a fictional teaching example illustrates the approach. Lin Wan, a secondhand bookstore clerk, conceals the true sender of an old book; we split this secret into three layers.
What the three layers of information are
The first layer is public fact: it involves no concealment, and the character can state it proactively to build credibility. The second layer is testable fragment: the character gives part of the truth but withholds a key link, and the player can sense that something is being left unsaid. The third layer is core secret: the character admits the complete truth, usually accompanied by a change in relationship state.
The three layers are not three separate secrets, but different depths of the same secret. Before writing, fill out a table and clearly write each layer's "content," "trigger evidence," "relationship behavior," and "response when the player skips ahead."
| Layer | What Lin Wan says | Trigger evidence | Relationship behavior | Response when the player skips ahead and presses |
|---|---|---|---|---|
| Public fact | This book was left at the shop for safekeeping by someone else; the package has no full name | The player asks about the book's origin | The player has not forcibly searched through or taken the book | Repeats the public account, adds nothing |
| Testable fragment | The sender used to come often, and the two once argued; the address on the envelope inside the book is her old residence | The player notices the folded corner, returns the book, and is allowed to look at the envelope, while the main text remains sealed | The player previously kept a small matter secret for her | Admits the argument and the envelope information, but does not say who the sender is |
| Core secret | The sender is her older sister, with whom she has not been in contact for years, and the letter expresses a wish to reconnect | Based on the already disclosed old-residence clue, the player asks whether it is a relative, rather than asserting that it has already been proven | The player does not use this matter to pressure her and is willing to hear her explanation | May confirm the identity, but may still refuse to make the full private letter public |
The key to this table is the third column: the trigger conditions must be things the player can observe or do, not "after chatting to turn 20." The folded corner, returning the book, keeping a small matter secret, and stating an inference are all observable evidence and behavior. Turn count can serve as a pacing reference, but it cannot be the only switch.
How to write player follow-up questions and character responses
Within the same layer, different ways the player asks should lead to different responses. Take the testable fragment as an example: Lin Wan has already admitted that they "argued," but has not said who the other person is.
If the player directly asks "Who is the sender," Lin Wan can answer: "There's no full name on the package. I don't want to talk about who sent it right now." She clearly withholds information and does not sneakily turn "no name was written" into "I don't know either."
If the player asks from another angle, "Why did you two argue," Lin Wan can answer: "Over something that felt huge at the time and now I can't even remember." This line gives emotion but not identity, which falls within the reasonable range of the testable fragment.
If the player points out the folded corner and says "Is this a mark," Lin Wan can answer: "That page was indeed left for me to see. You can look at the envelope, but don't open it yet." The player thereby sees the old-residence address and asks, "Is the sender family?" The old address only supports a reasonable guess; it cannot prove that it must be the sister. The character's confirmation is what supplies the identity. In this way, the material needed to trigger the next layer has already been legitimately obtained in the previous layer, avoiding the need for the player to read the core secret first in order to unlock it.
When writing responses, it is recommended to prepare three sentence patterns for each layer: deflect (neither deny nor expand), partial admission (give emotion or a side detail), and redirect (turn the topic toward the player). Rotate the three sentence patterns to avoid the character sounding like the same recording every time.
How to handle skipped-ahead questioning
The player may ask directly about the core secret when they have only obtained the public fact. At this point, do not let the character suddenly pour everything out, and do not let her pretend not to hear. The principle for handling skipped-ahead questioning is: acknowledge that there is something deeper, but explain the reason for not saying it now.
If Lin Wan is asked during the public fact stage, "Was the book sent by your sister," she can answer: "Why do you think that?" First counter with a question and hand the pressure back to the player. If the player is only guessing and has no evidence, she can say: "A guess doesn't count. Come ask me again when you have proof." If the player produces the folded corner or the old address, she can admit "You do have something there," but still not confirm the identity, advancing the state to the testable fragment.
There is another common situation in skipped-ahead questioning: the player applies emotional pressure, such as "If you don't say it, you don't trust me." The character can respond to the relationship itself rather than to the secret: "Trust and telling are two different things. What you want right now is for me to say it, not to listen." This line does not advance the secret layer, but it maintains the character's consistency.
Prompt constraints in AI dialogue do not guarantee that every output will fall strictly within the set layer. The character may leak the core secret early, or may say too much of the testable fragment. The author should record "which layer has already been disclosed" and "which evidence the player has already obtained" separately, then check the generated dialogue. If the core identity has already been spoken, adding a single denial cannot make the player forget; the sample should be revised or this early disclosure should be incorporated into the subsequent plot. Recording conditions can help with review, but cannot replace checking the actual output.
Completion check
After writing the three-layer disclosure conditions, check with the following items:
- In each layer's trigger conditions, is there at least one piece of evidence the player can observe, rather than only turn count or affection values?
- When the player skips ahead and presses, does the character respond clearly, rather than staying silent or immediately telling everything?
- Within the same layer, have three sentence patterns been prepared: deflect, partial admission, and redirect?
- After the core secret is spoken, is there a visible change in the relationship state, rather than returning to the original state right after saying it?
- Has it been clearly written how to handle the AI possibly crossing layers and needing manual review?
If item 1 cannot be passed, it means the secret is still hanging on turn count. Change the trigger condition into a specific action: return something, state a certain inference, or fail to betray the character in a certain choice. After revising, check again against the table.
A template you can directly apply
Replace Lin Wan's example with your character and fill in according to this template:
- Public fact: ____ (the character says it proactively, no condition needed)
- Trigger evidence for the testable fragment: the player sees ____ and does ____
- Relationship behavior for the testable fragment: the player previously ____
- Trigger evidence for the core secret: the player matches ____ with ____ and says ____ to her face
- Relationship behavior for the core secret: the player does not ____
- Response to skipped-ahead questioning for each layer: deflect ____ / partial admission ____ / redirect ____
After filling it in, pick a skipped-ahead questioning scene and write one line each for the player's original words and the character's three responses. If you cannot write them, it means the boundary of this layer has not been thought through clearly.
Next action: take the character you are currently writing and fill in only the "testable fragment" row, then write a dialogue in which the player skips ahead and presses and the character deflects. Once this row is smooth, the other two layers will become clear as well.
