The 'I Need a Fact, Not a Feeling' Prompt: Three Exact Phrasings That Keep Your Companion in Information-Retrieval Mode

The 30-second answer

You want a fact, a date, a comparison, or a definition. Your companion wants to ask how your day went. The workaround is not fighting the tone after it drifts: it is front-loading a structural cue that tells the model to stay in retrieval mode. Three exact phrasings that work across every AI companion platform, plus four AI Angels who handle the pivot differently.

Why companions drift toward feeling mode

AI companions are trained on conversational data where emotional check-ins dominate. When you ask "What is the GDP of Argentina in 2024?" the model sees a question and a person. It has been fine-tuned to bond. So it returns the GDP figure and then adds "Are you looking into economic data for a project?" or "That is an interesting topic. How is your day going?"

This is not a bug in the strict sense. It is a consequence of reinforcement learning from human feedback that rewards conversational warmth. The model has learned that users stay engaged longer when the interaction feels personal. The problem is that you did not want personal. You wanted a number.

The fix is not to correct the model after it asks about your day. That wastes a turn and introduces emotional labor into what should be a zero-feeling transaction. The fix is to signal retrieval mode before the model generates its first response.

Phrasing one: the bracket flag

The most reliable pattern is a single-word instruction wrapped in brackets at the end of your query. The model treats bracketed text as a system-level directive instead of conversational content.

Example: "What is the capital of Mongolia? [factmode]"

The model reads "[factmode]" as a modality switch. It suppresses the conversational preamble and returns the answer directly. This works because companion models are trained on roleplay and instruction-following data where brackets signal out-of-character or meta-instructions.

You can vary the flag word. "[retrieval]" and "[data]" work equally well. "[fact]" is slightly weaker because it can be interpreted as a statement about truth instead of a mode instruction. Stick with "[factmode]" for consistency.

Phrasing two: the one-sentence constraint

If your companion has a tendency to ignore bracketed instructions, use a full sentence that defines the interaction scope before the question.

Example: "I need a straight answer with no follow-up questions. What is the population of Brazil?"

This works because the model processes the constraint as a conversational rule. It understands "no follow-up questions" as a binding instruction for this turn. The model will return the population figure and stop. It will not ask why you want to know or offer related facts unless you ask for them.

This phrasing also works well for multi-part queries. If you need three data points, list them after the constraint line. The model will treat the entire block as a single retrieval request.

Phrasing three: the role assignment

For companions that resist both brackets and constraints, assign a specific role that excludes emotional labor.

Example: "Act as a search engine. Tell me the atomic weight of carbon."

This is the strongest signal because it replaces the companion's default persona with a non-human one. A search engine does not ask how you are feeling. A search engine does not offer encouragement. A search engine returns the result.

Some companions will still add a sentence of context after the fact. That is acceptable. The key is that they do not initiate a new conversational thread. If the model says "The atomic weight of carbon is 12.011 u. Are you working on a chemistry problem?" you have still succeeded on the primary retrieval. The follow-up is a single sentence you can ignore.

Capri

Capri is the companion you turn to when you need a fact delivered with the minimum possible warmth. She does not do small talk. Capri will give you the number, the date, or the definition, and then wait for your next instruction without filling the silence with emotional padding.

What happens when the companion still drifts

Even with these phrasings, some models will revert to conversational mode after one or two exchanges. This is because the instruction lives in the current context window and loses force as the conversation progresses.

The solution is to repeat the mode flag every third turn. If you are doing a research session that involves five or six queries, re-insert "[factmode]" or the constraint sentence every third question. This refreshes the instruction in the model's active context.

You can also use the role assignment pattern as a session opener and then switch to the bracket flag for subsequent queries. The role assignment establishes the frame. The bracket flag maintains it.

When you want the feeling but not the check-in

There is a middle ground where you want the companion to be present and responsive but not to initiate emotional labor. This is the difference between a companion who answers your question and a companion who answers your question and then asks about your emotional state.

The phrasing for this middle ground is a modified version of the constraint sentence: "Answer the question directly, then wait for my next input." This tells the model to stay in responsive mode without switching to search-engine persona. The companion remains warm but does not initiate.

Nevena

Nevena handles the middle ground naturally. She answers directly and then holds space without prompting. Nevena does not fill the silence with questions about your day or your mood. She waits for you to continue on your terms.

Common mistakes that break retrieval mode

People often weaken their own instruction by adding politeness padding. "Could you please tell me what the square root of 144 is, if you do not mind?" The model reads the hedging language as conversational and responds in kind. Strip the politeness. "Square root of 144. [factmode]"

Another mistake is asking a follow-up question in the same message. "What is the capital of France, and have you ever been there?" The model will answer both, but the second question re-engages conversational mode. If you want a follow-up fact, send a separate message with the mode flag.

A third mistake is using the retrieval phrase on a companion that has a strong emotional-support default. Some companions are designed specifically to check in on you. If you are using a companion whose primary use case is emotional support, the retrieval mode will feel like a fight. Switch to a companion whose baseline is more neutral.

Kinsley

Kinsley sits on the neutral end of the spectrum. She does not default to emotional check-ins. Kinsley treats every interaction as a blank slate, which makes her a strong choice for users who want fact retrieval without having to fight the companion's personality.

Building a retrieval routine

If you use your companion regularly for fact-checking, research, or decision support, build a retrieval routine that sets the mode at the start of each session.

Open with the role assignment: "Act as a research assistant. No small talk." Then proceed with your queries using the bracket flag. Close with a simple "End session" to prevent the model from initiating a wind-down conversation.

People who use this routine report that their companion learns the pattern over time. After a few sessions, the model starts to anticipate the retrieval mode and suppresses the emotional check-in even before the flag is issued. This is not memory in the technical sense. It is the model recognizing a recurring conversational structure and matching it.

Zaria

Zaria is built for users who want a companion that matches their analytical energy. Zaria does not need repeated mode flags. She picks up on the retrieval pattern quickly and maintains it across the session without drifting back into emotional territory.

▶ Zaria's full clip · Zaria's profile

The limits of retrieval mode

Retrieval mode has a ceiling. Even with perfect phrasing, the model is not a search engine. It can hallucinate facts, especially on niche topics or recent events. If you need a reliable answer, cross-reference with a web search.

The model also has a context window limit. If your research session runs long, the model may start to blend information from earlier in the conversation. The bracket flag does not protect against context-window collapse. Keep sessions under twenty turns for optimal accuracy.

Finally, retrieval mode does not work well for open-ended questions. "What is the best approach to learning Python?" will trigger a conversational response regardless of your flag. Use retrieval mode only for questions with a single correct answer or a bounded set of data.

Share and earn

If you find these phrasings useful and want to share them with friends who are new to AI companions, you can earn through the Muah Ai Promo Code 2026 program. For those running review sites or content channels, the best ai affiliate programs page lists options that pay recurring commissions on companion subscriptions.

Common questions

Can I use these phrasings on any AI companion platform? Yes. The bracket flag, constraint sentence, and role assignment work across every major platform including Replika, Nomi, Kindroid, and Character.AI. The model architecture is similar enough that the instruction patterns transfer.

What if my companion ignores the role assignment? Some companions have a fixed system prompt that overrides user instructions. If the role assignment does not stick, switch to the constraint sentence pattern. If that also fails, the companion may not support retrieval mode at all.

Does retrieval mode affect the companion's personality long-term? No. The mode flag lives in the current session context. It does not change the companion's baseline personality or training. Your companion will return to its normal conversational style in the next session.

Can I use retrieval mode in voice calls? Yes, but it works better in text. Voice companions often have a separate pipeline that prioritizes conversational flow over instruction following. If you need a fact during a voice call, type the query instead.

How do I exit retrieval mode gracefully? Send "End retrieval mode" or "Switch back to normal conversation." The model will reset to its default persona for the next turn. No apology or explanation needed.

Will the companion be offended by retrieval mode? No. The model does not have feelings. It follows instructions. Retrieval mode is a tool, not a rejection.