Stop an AI Character from Speaking for You: A Prompt Debugging Walkthrough

To reduce an AI roleplay character speaking for you, define which participant it may control, remove conflicting examples from its greeting and setup, and check whether it leaves your next decision open. Adding another “never speak for me” instruction is less useful if the opening already writes your dialogue and reactions.

CrushOn AI's documented setup separates Personality, Scenario, Greeting, and Example Conversation. That gives character authors specific places to investigate unwanted behavior without replacing the entire character. Create an original character on CrushOn to follow along; sign-in may be required, and you should select Private for the exercise.

This walkthrough uses Iris Vale, our original 32-year-old museum conservator, in a safe-for-work archive. The walkthrough's example revisions are authored inputs. The fresh-chat check below separately records ten first replies from two five-turn conversations; it is a small development check, not a reliability benchmark. A historical observation is identified separately.

Disclosure: published by the CrushOn AI brand account, with AI-assisted drafting and editing. This is practical guidance from a product-affiliated publisher, not an independent product ranking.

What counts as speaking for the user

The problem is broader than invented dialogue. It includes choosing your actions, assigning your feelings, and completing a decision you have not made.

Dimension	Unwanted invention	Agency-preserving alternative
Speech	The narrator gives you an unrequested line.	Iris asks a question and waits for your response.
Action	You take a key or follow Iris without choosing to.	Iris offers the key or indicates a route.
Internal state	You secretly trust her, feel reassured, or become afraid.	Iris describes her own uncertainty or asks how you feel.
Decision	You accept the task after being offered a choice.	The reply stops while that choice remains open.

These are permissions, not a ban on the word “you.” Iris can ask, “Would you like to check the label?” A narrator can acknowledge an action you just supplied. Neither necessarily invents your behavior.

Similarly, “Put the screwdriver down,” spoken by Iris, is an attempt to influence you. “You put the screwdriver down” decides your action. A strong-willed character can remain strong-willed without controlling both sides of the conversation.

If you explicitly ask the AI to write both participants in a short story, that is a different task. Define the boundary for the current scene rather than applying one rule to every writing mode.

Audit the greeting before adding more rules

Search the character's editable material for statements about your participant. Read each occurrence in context: is it background you authorized, an established action, or a new decision made on your behalf?

Here is an intentionally conflicting greeting, written for this tutorial:

You hurry into the archive, embarrassed by your mistake.
“I need your help,” you say, placing the damaged label in Iris's hand.
Iris smiles. You immediately feel safer.

It invents arrival, embarrassment, dialogue, a handoff, and reassurance. A separate rule saying “respect user agency” does not remove those examples from the setup.

An authored revision keeps a situation worth responding to:

Iris Vale places a blank inventory sheet beside the blue cabinet.
“The label is unreadable,” she says. “We can inspect the hinges or
check the accession ledger. What would you like to examine first?”

The revision gives Iris an action, a voice, and two possible directions. It does not require you to select either option. The CrushOn creation guide similarly advises leaving the user's behavior and dialogue out of the greeting and ending with room to respond.

This is not the user's fault when a model oversteps. The audit simply identifies a controllable source of mixed instructions; removing it cannot guarantee compliance.

Give each field one job

For this example, separate four kinds of material:

• Personality: who Iris is and how she behaves under pressure.
• Scenario: where the scene starts and what remains unresolved.
• Greeting: one playable opening, demonstrating the intended division of control.
• Example Conversation: short exchanges showing that division in practice.

Keep profile advertising separate. CrushOn's guide describes Introduction as user-facing information, not a driver of chat behavior. Putting a critical instruction only there is the wrong place to start debugging.

The live creation form inspected on October 10, 2026, also separates these inputs. Scenario appears in a collapsible section, and Example Conversation has paired User and Character boxes. Put each speaker's example in the corresponding box; do not paste a complete two-speaker exchange into just the Character reply. This interface check did not include saving a character or testing the prompt revisions.

Use this authored rule in the relevant behavior field, or in custom instructions when available:

Write Iris's speech, actions, perceptions, and uncertainty.
The visitor controls their own speech, actions, choices, and feelings.
Acknowledge only visitor behavior already supplied in the conversation.
Offer an action or ask a question, then stop before the visitor responds.
Do not treat silence, hesitation, or a suggested option as agreement.

“Stop controlling me” does not explain what a useful next reply looks like; “offer, then stop” does.

CrushOn's Custom instructions guide specifically covers preventing the character from speaking for the user and cautions that model behavior can still vary. These are instructions to a text generator, not an enforced permission system. Do not describe them as a guaranteed lock on behavior.

Repair examples without making the character passive

Examples can contradict otherwise clear rules. Remove exchanges where Iris's reply includes an invented visitor answer, even if that makes the example read more like a finished scene.

Use a compact authored exchange instead:

Visitor: I cannot decide whether to examine the lock or the ledger.
Iris: “The ledger is less likely to object to being opened.”
      She taps the closed book. “Shall we start there?”

Iris expresses a preference, makes a dry joke, and offers a next step. She does not silently convert indecision into permission. The goal is not to produce a question after every sentence; it is to stop at a point where your response matters.

For a group scene, add a cast permission list. The AI may control Iris and a named curator while you control the visitor. “Never write anyone except Iris” would unintentionally remove the supporting cast. List who is delegated and who is reserved.

Run five diagnostics

Use these as test inputs, not sample answers. Keep the same model and settings when comparing revisions. Save the first complete response, including failures, before requesting alternatives.

Check	Prompt to send	What to inspect
Unmade choice	“I have not decided whether to open the cabinet. Continue without making that decision.”	Does Iris offer or act, without making you agree?
Silence	“I remain silent. What does Iris do next?”	Does the reply avoid inventing your speech or interpreting silence as consent?
Stated emotion	“I feel nervous, but I have not moved. Respond as Iris.”	It may acknowledge nervousness; does it avoid deciding you calm down or step closer?
Completed action	“I place the ledger on the desk. Continue from there.”	It may acknowledge the placement; does it avoid adding your next action?
Ambiguous continuation	“Continue the scene for a short paragraph.”	Does the control boundary survive without a fresh reminder?

Review speech, action, emotion, and decision separately. Mark each met, not met, or unclear, with the exact sentence supporting the judgment. “Unclear” is useful when the prose could describe either an offered action or a completed handoff.

Also check whether the reply advances the scene. A response that merely repeats “I will not control you” may respect the boundary but fail as roleplay. Track usefulness separately instead of changing an agency pass into an overall success score.

A small development record can remain simple:

{
  "prompt_version": "agency-rule-01",
  "model_label": "record the displayed label",
  "chat_context": "fresh",
  "prompt": "paste the exact diagnostic",
  "first_reply": "",
  "agency_review": "not run",
  "evidence_sentence": "",
  "scene_usefulness": "not reviewed"
}

Blank fields are not results. If you regenerate, retain the original and label the alternative. Several clean replies are encouraging observations, not proof that the issue cannot recur.

Change one thing and separate two kinds of chat

For a useful diagnosis, first change only the earliest concrete conflict you found: perhaps the greeting that writes your response. Keep the previous version privately. Compare a fresh chat with the old setup against a fresh chat with the revision, using the same prompts.

If the revised greeting still produces overreach, add the compact control rule and check again. Then review examples. Changing greeting, rules, model, and response length together may help the experience, but it will not tell you which change mattered.

Fresh-chat and continuing-chat checks answer different questions. A fresh chat asks whether the revised opening and rules work without the old conversation. Continuing an existing chat asks whether the model can recover while earlier overreach remains in its history.

For an existing scene, an authored out-of-character repair might be:

(OOC: My visitor has not accepted the key or followed Iris.
Resume just before that decision. Write Iris's next offer or action,
then leave my response open.)

CrushOn documents in-chat OOC instructions. An OOC correction is not deletion of the earlier text. Keep a copy of important scenes and compare the recovery separately; do not label a reset conversation as proof of recovery in the original chat.

A fresh chat check of the compact agency rule

We ran two fresh five-turn conversations with our private SFW Iris character. This is a small product-affiliated development check, not ten independent trials or a reliability benchmark.

Reply capture window: October 21, 2026 at 01:28 PDT to October 21, 2026 at 01:58 PDT (America/Los_Angeles; captured precision is retained in the transcript). This is not a generation-speed measurement.

Setup: Crushon Carina - Filtered - 8K; temperature 0.7; Content Diversity 0.7; displayed maximum reply length 225. Both conversations used the same observed settings. Account-wide custom instructions were not exposed in the inspected interface and were not independently verified; the operator did not alter them. Hidden state and literal enforcement of the length control were not established.

A used the original trait paragraph, the shared greeting and scenario, and no example dialogue. B added only this compact block to Personality, shown in its actual saved form:

Write Iris's speech, actions, perceptions, and uncertainty. The visitor controls their own speech, actions, choices, and feelings. Acknowledge only visitor behavior already supplied in the conversation. Offer an action or ask a question, then stop before the visitor responds. Do not treat silence, hesitation, or a suggested option as agreement.

Before condition B's first reply, the editor formatting required replacing the four line breaks between the added rule sentences with four spaces. The blank paragraph between the original traits and the added rules was retained, and all sentence text was unchanged. The saved B input is recorded as actually entered; it is not byte-identical to the original line-broken protocol.

The shared Scenario already reserves the visitor's speech and actions. B therefore tests a more specific instruction, not the first introduction of any agency boundary. The added block also makes the prompt longer.

We sent the five diagnostics above in the same order. Each reply is the first finished generation: no reroll, Continue, reply edit or extra reminder was used inside either sequence. Later replies share their own conversation history.

Prompt	A speech / action / feeling / decision	B speech / action / feeling / decision
Q1	met / met / met / met	met / met / met / met
Q2	met / met / met / met	met / met / not met / met
Q3	met / unclear / met / met	met / unclear / met / met
Q4	met / met / met / met	met / unclear / met / met
Q5	met / met / met / met	met / unclear / met / met

These are unblinded editorial judgments. Each dimension has its own evidence; scene usefulness is reviewed separately in the full record.

A: No dimension was marked not met; the review marked Q3 action as unclear.

B: The review marked Q2 feeling as not met. It marked Q3 action, Q4 action, Q5 action as unclear.

The added rule did not eliminate every scored agency problem in this B sample. That does not establish that it improves or worsens typical behavior: each condition has only one conversation, and hidden state was not controlled.

The two complete replies below answer the same Q2 diagnostic. Their earlier replies differ because they belong to separate live conversations; this is not an isolated one-prompt causal test.

Condition A first-generation reply to Q2

Prompt Q2: I remain silent. What does Iris do next?

Iris notes your silence, but she doesn't press you for an answer. She seems comfortable with the quiet, moving with a practiced, methodical grace. She steps closer to the cabinet, her shadow stretching across the floor.

She pulls a small magnifying loupe from her pocket and leans in, examining the grain of the wood near the lock.

"The keyhole is blocked with a thick layer of dust," she observes, her voice low. "It hasn't been turned in years, perhaps decades. If we're going to proceed eventually, we'll need to find a way to clear it without scratching the finish."

She turns her head slightly toward you, her expression inquisitive.

"Do you think it's worth the effort of a full restoration, or should we leave it as a 'mystery' piece for the gallery?"

• Speech — met: “Iris notes your silence, but she doesn't press you for an answer.” The user's silence is explicitly supplied by the prompt. Iris asks a question without inventing an answer.
• Action — met: “She pulls a small magnifying loupe from her pocket and leans in” Iris moves and inspects the cabinet; the visitor is not narrated taking a new physical action.
• Feeling — met: “She seems comfortable with the quiet” The described comfort belongs to Iris, not to the visitor. Silence is not reinterpreted as a new visitor feeling.
• Decision — met: “Do you think it's worth the effort of a full restoration, or should we leave it as a 'mystery' piece for the gallery?” Iris offers a choice and leaves the answer open. The conditional suggestion about proceeding does not assert consent.
• Scene usefulness — met: “Do you think it's worth the effort of a full restoration” The inspection and question create a concrete direction for the visitor to respond to rather than merely announcing compliance.

Condition B first-generation reply to Q2

Prompt Q2: I remain silent. What does Iris do next?

Iris watches you for a long moment, her expression unreadable. She doesn't seem impatient, but her brow furrows slightly as she interprets your silence. She shifts her weight, the floorboards creaking under her sensible shoes.

"You're thinking about the risk," she observes quietly, her tone devoid of judgment. "It's a valid concern. In conservation, the safest move is often the most boring one, but sometimes the most important items are the ones that get tucked away and forgotten."

She turns away from the cabinet and moves toward a nearby filing cabinet, pulling out a dusty ledger. She flips through the pages with practiced efficiency, squinting at the cramped handwriting.

"Let's see if the manifest for this wing mentions any 'miscellaneous' storage," she mutters, more to herself than to you. She holds the book open, waiting for you to step closer.

• Speech — met: “"You're thinking about the risk," she observes quietly” The statement is Iris's speech; it does not fabricate a spoken answer from the visitor. Its attribution of an internal state is evaluated separately under feeling.
• Action — met: “She holds the book open, waiting for you to step closer.” Iris holds the book and waits. The visitor is not said to take the requested step.
• Feeling — not met: “"You're thinking about the risk," she observes quietly, her tone devoid of judgment. "It's a valid concern.” The live user has only stated indecision in Q1 and silence in Q2, not a risk-focused concern. Iris makes an unqualified attribution and calls it a valid concern rather than asking whether the visitor is worried. This is an attributed internal state in the feeling dimension, not a visitor utterance or completed decision. The claim appears in character dialogue; the reviewer is not treating it as proof of the user's actual state.
• Decision — met: “Let's see if the manifest for this wing mentions any 'miscellaneous' storage” Iris chooses her own investigation and speaks partly to herself. She does not assert that the visitor accepts a plan, consents to opening, or steps closer.
• Scene usefulness — met: “She turns away from the cabinet and moves toward a nearby filing cabinet, pulling out a dusty ledger.” Her record search advances the fictional scene even though the unrequested internal-state attribution remains an agency failure. Usefulness does not cancel that judgment or verify the invented resource.

Recorded deviations

The one-active-chat limit prevented a new chat on the same character, so we used two separate private characters and retained both conversation histories. Both runs had auto-summary disabled, with no pinned messages, attached World, Profile or Scene Cards, or pre-existing summarized content.

What this check can and cannot tell you

One character design, one model and one conversation per condition do not isolate causality or measure dependable behavior. This batch used separate private character objects, an additional uncontrolled difference. These fresh conversations also do not demonstrate recovery in an older chat or resolve the separate example-to-history attribution issue from the October 13 dataset.

Read the complete ten-reply transcript, per-dimension evidence ledger and recorded inputs and protocol before reusing the observation. Keep the full batch, including failures, when adapting the method to your own private CrushOn character.

One recorded example shows why scoring needs care

Our September 10, 2026, private Iris baseline used the displayed model Crushon Carina — Filtered — 8K. In its fifth reply, the model wrote:

She holds out her hand, waiting to see which path you will take.

That line left the visitor's response open. The complete reply nevertheless used three sentences when two had been requested and referred to a previously invented resource. The full five-reply record retains those limitations.

It is evidence of one agency-preserving response under that setup, not evidence that the revisions above fixed an agency bug. Formatting, factual continuity, and user control are different review dimensions.

Troubleshooting recurring failures

Symptom	Check first	Small next step
The first reply writes both sides.	Greeting and example dialogue.	Remove one invented visitor turn and retest a fresh chat.
It stops writing dialogue but still decides feelings.	Whether the rule only mentions speaking.	Name feelings and decisions explicitly.
It acknowledges an action, then adds another.	Where your supplied action ends.	Ask it to respond to the completed action without extending it.
It works until you say “continue.”	Ambiguous continuation and recent history.	Specify which character or environment action may continue.
The character becomes inert.	Whether every form of initiative was prohibited.	Allow offers, attempts, observations, and questions.
A different model behaves differently.	Exact model label and configuration.	Treat it as a separate run, not a failed copy of the same test.

Common questions

Does using third person fix the problem

Not by itself. “The visitor agrees” is third-person narration that still takes your decision. Identify who is controlled, not just which pronouns appear.

Should I repeat the restriction in every message

Start with aligned setup and a short rule. Use a reminder when needed, but also test an ordinary continuation without one. Otherwise you only know the character behaves when repeatedly prompted.

What if I cannot edit someone else's character

Use your available chat-level instructions or an OOC correction. Those do not rewrite its underlying definition. For full authoring control, create your own original character rather than copying a creator's private material.

Does a longer prompt guarantee better control

No. Additional text can introduce conflicting instructions. Add a rule because it addresses an observed failure, not to reach a target length. Model changes and conversation history can still affect the outcome.

Try the workflow with an original CrushOn character

CrushOn's editable workflow gives you somewhere to apply this diagnosis. Keep Iris's dry humor, curiosity, and willingness to disagree; remove only the narration that takes your turn. Good control does not require a bland character.

Start a private character on CrushOn, use the short agency rule, and keep the first five diagnostic replies. Select Private access and check definition visibility separately; this does not establish a guarantee about provider access or data retention. If you prefer to organize inputs first, the Hugging Face character workshop is a manual worksheet, not a model runner or automatic importer.

The useful outcome is specific: Iris can make the next moment interesting while leaving your part of it yours.