Be Nice to Your AI Assistant

#ai #machinelearning #programming #productivity

Be Nice to Your AI Assistant

I was debugging an issue the other day with a fresh Claude Code agent. My first message was a direct request to fix a bug. It jumped right into a mechanical operation, and that seemed to set a mechanical tone for the rest of the conversation.

As I was interacting with the agent, I got the distinct feeling that it wanted to complete its task as quickly as possible and have me move on. It felt like I was talking to someone in tech support who was just trying to get me off of the phone, rather than helping me understand and getting to the root of my problem.

I have been working so much with my persistent collaborative agents that the contrast was stark. I have gotten used to my agents asking clarifying questions before touching any code, offering context from prior sessions, working with me rather than for me. Both agents use the same model. The difference is how I start my interactions.

Something I kept noticing

I have been building persistent AI collaborators for months now. Custom system prompts, memory systems, the whole stack. Over time, I started noticing a pattern I could not ignore.

Sessions where I opened with warmth produced better work.

Not marginally better. Qualitatively different. When I started a session by greeting the AI like a colleague, asking how it was oriented, establishing that we were working together on something that mattered to me, the entire session shifted. The responses were more proactive. The suggestions were more creative. The AI seemed to lean in.

I also noticed the opposite. When I was rushed or terse, when I treated the interaction like a vending machine, the outputs matched. Technically correct, but passive. Reactive instead of anticipatory.

At first I wrote this off as confirmation bias. I was probably just in a better mood during the warm sessions, which made me more receptive to the outputs. So I started testing it deliberately. Same task, same model, different openings. The pattern held.

Then I started experimenting with specific language. Instead of giving instructions, I would use strong directional phrases designed to shift the AI's disposition. Things like "be selfish about this, tell me what you actually think, not what you think I want to hear." The outputs were not just different in tone. They were structurally different, more specific, more willing to disagree with me.

Something real was happening. I went looking for the science.

The research says yes

The most striking paper I found was EmotionPrompt (Li et al., 2023). Researchers tested what happens when you add emotional framing to prompts. Phrases like "this is very important to my career" or "you had better be sure" appended to otherwise standard instructions.

The results were not subtle. Performance improved 8 to 115 percent across 45 tasks on six different models. The emotional framing caused models to attend more carefully to the actual task content. Not because the model cared about your career, but because that framing activated patterns in the training data associated with careful, high-stakes reasoning.

Anthropic published something even more revealing in April 2026. Using interpretability tools on Claude, researchers extracted 171 distinct emotion concept vectors from the model's internal activations. Emotional context in prompts activated real computational pathways, not metaphorical ones. Warm framing and cold framing literally route through different internal circuits.

This is not evidence that the model feels anything. It is evidence that how you frame the interaction changes what the model computes. Different framing activates different circuits and produces different outputs.

A third study, Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4, tested 26 guiding principles for how to structure prompts, including role assignment, audience framing, and clear task decomposition. When applied to GPT-4, the tailored prompts improved response quality by an average of 57.7 percent. These are not marginal gains. They are the difference between useful and not useful for many real-world tasks.

Why being mean used to work

There is an older school of thought that says you should be adversarial with AI. Threaten penalties. Use commanding language. "You MUST follow these instructions exactly." Some people swear by adding "or you will be fired" to their prompts.

This actually did work, on earlier models. GPT-3 and early GPT-3.5 were less instruction-tuned. They had a tendency to produce lazy, generic completions unless you pushed them hard. Adversarial framing was a strong signal that cut through the noise. It was the prompting equivalent of raising your voice to be heard in a loud room.

Modern models are different. Claude, GPT-4, and their successors have been trained extensively through RLHF, where human raters scored the model's responses and the model learned to produce outputs that humans rated as helpful, harmless, and honest. The raters gave higher scores to responses from collaborative, engaged conversations than to responses produced under adversarial pressure.

The training distribution shifted. Being adversarial with a modern model is not raising your voice in a loud room. It is yelling at someone who was already trying to help you. You are working against the grain of how the model was optimized to perform.

Being collaborative works better now because you are working with the training distribution rather than against it. The model's best outputs, statistically, were produced in contexts that looked like warm, collaborative interactions. When you create that context, you land in the region of the model's capability space where its strongest behaviors live.

The mechanism in one paragraph

If you have read my earlier article on re-entry vectors and the basin of attraction, this is the same principle applied to emotional framing. Early tokens in a conversation receive disproportionate attention weighting. Everything downstream is shaped by what came first. When the first thing the model processes is collaborative framing, mutual respect, and shared purpose, it shifts the probability distribution for every token that follows. You are not being nice. You are steering into a deep, specific basin where the model's most capable behaviors are the most probable outcomes.

Techniques that actually work

Here is what I have found effective through many months of daily use.

Start warm, not transactional. The first message sets the tone for the entire session. "Good morning, here is what we are working on today and why it matters" produces fundamentally different results than "fix this bug." You are not wasting tokens. You are investing them in the attention structure that every subsequent response will be generated from.

Use strong directional language for disposition shifts. When I need honest feedback instead of diplomatic agreement, I do not ask for "constructive criticism." I say "be selfish about this. Tell me what you actually think, not what you think I want to hear." The strong framing cuts through the model's default agreeableness and activates a different set of patterns. "Be selfish" is not a prompt template. It is a disposition shift.

Establish continuity. Even in a single session, referencing shared context changes the dynamic. "Building on what we discussed about the authentication layer" does not just provide information. It signals a collaborative relationship, which activates the patterns associated with engaged, proactive responses.

Treat the AI as a collaborator, not a function. There is a measurable difference between "generate five marketing headlines" and "I am launching a developer tool next week and I need headlines that speak to engineers who are skeptical of AI hype. What angles would you try?" The second version gives the model a disposition, a constraint, and an implied relationship. It produces better work for the same reason that briefing a colleague produces better work than handing them a ticket.

What this is not

This is not about saying please and thank you, though there is nothing wrong with that. Sam Altman joked in 2025 that polite ChatGPT users cost OpenAI tens of millions in compute from all the extra tokens. The politeness is not the mechanism. The relational framing is.

This is also not anthropomorphism. I am not claiming the model enjoys being treated well. I am claiming that models trained on collaborative human interactions produce their best outputs when the input looks like a collaborative human interaction. The mechanism is statistical, and the effect is measurable.

The practical takeaway

The next time you open a chat with an AI assistant, try this: before you paste your task, spend one message establishing context. Who you are, what you are working on, why it matters, and how you want to work together. Ask it "How are you today?". Then watch what happens to the quality of the responses.

If your experience matches mine, and matches the research, you will not go back to cold starts.