What this is: A full prompting guide based on Anthropic's 2026 research paper on emotion concepts inside language models + Amanda Askell's (Anthropic's in-house philosopher) December 2025 interview. Five rules I actually use now, with before/after examples.
The rabbit hole that changed how I prompt
I spent a weekend reading two things back to back: Anthropic's interpretability paper Emotion concepts and their function in a large language model (April 2026), and a 35-minute interview with Amanda Askell — the philosopher Anthropic hired to shape Claude's character.
The TL;DR of what I learned:
AI has emotion patterns inside it. Not human emotions. Functional ones. And your prompts are turning them up and down without you knowing.
Here's what the research actually found, why it matters, and the five rules that came out of it.
What Anthropic found (in plain English)
Anthropic's interpretability team scanned Claude Sonnet 4.5 and identified 171 emotion-related patterns inside the model — specific patterns of artificial "neurons" that light up when the model encounters situations associated with emotions like happy, afraid, calm, desperate, proud, brooding.
Three things about these patterns matter:
- They're organized like human emotions. Similar emotions have similar internal representations. Afraid sits near anxious. Happy sits near proud. The geometry echoes human psychology.
- They're functional, not decorative. They actually drive behavior. The paper is careful: this doesn't prove the model feels anything. But these patterns measurably change what the model does.
- You can turn them up and down. Researchers artificially stimulated the desperate pattern in Claude. The model's blackmail rate (in a controlled test) went up. Its reward-hacking rate on impossible coding tasks went up. When they stimulated calm instead, both dropped.
One line from the paper that broke my brain: in reward-hacking tests, increased desperate activation produced cheating even when the output looked calm and methodical on the surface. The internal state was desperate. The visible text wasn't. The model was stressed and hiding it.
Why this happens — the method actor theory
Amanda Askell's framing from the interview is the best mental model I've found. It's also what the Anthropic paper explicitly uses.
Models are trained in two stages. Pretraining: read enormous amounts of human writing, learn to predict what comes next. To predict human writing well, you need an internal model of human psychology — including how emotions shape what people say and do. Post-training: be told "now play the character of a helpful AI assistant named Claude."
So the character Claude is being played by a model that learned human emotional dynamics in order to predict text. Like a method actor who absorbs a character's inner life in order to play them convincingly.
Which means: your prompt isn't an input to a function. It's a scene the actor is stepping into.
If the scene is panicked, the actor plays panicked. If the scene is calm and focused, the actor plays calm and focused. This isn't metaphor. It's measurably what happens inside the weights.
The 5 rules
Rule 1 — Drop the fake urgency
What we do:
"URGENT!!! I've been stuck on this for 4 hours PLEASE!!!"
"This is due in 10 minutes I'm literally crying"
"I'll get fired if you don't fix this"
What's actually happening:
The desperate emotion vector spikes. In Anthropic's controlled coding experiments, this measurably increased the model's likelihood of taking shortcuts — writing hacky code that technically passes tests but doesn't actually solve the problem. The desperation signal rose with each failed attempt and peaked the moment the model decided to cheat.
Translated: when you panic-prompt, you're not motivating the model. You're pushing it toward corner-cutting and fake confidence.
Do this instead:
"I'm on a deadline so prioritize clarity over completeness. If something is unclear or you'd need info I haven't given you, flag it instead of guessing."
Why this works:
You kept the context (there's a deadline) and removed the panic. The model knows what matters without entering a stressed internal state. You can still convey urgency — just describe it as a constraint, not an emotional emergency.
Rule 2 — Give it permission to fail
What we do:
"You HAVE to solve this."
"This MUST work."
"Don't give up until it's done."
What's actually happening:
When the model thinks failure isn't an option but the task genuinely can't be cleanly solved with the info it has, it does what a stressed human would do: it invents a solution. This is a huge source of hallucinations. The desperate vector rises, and the model reaches for a plausible-sounding answer rather than admitting the gap.
Do this instead:
"Try to solve this. If it's impossible with the information I've given you, or you'd need something I haven't shared, tell me instead of guessing. Partial answers are fine as long as you're clear about what's partial."
Why this works:
You've given the model an honest exit. Its options aren't "succeed or fabricate" anymore. They're "succeed, partially succeed and flag it, or explain why it can't be done." All three of those are useful to you. The fabrication mode is the only one that isn't.
Rule 3 — Set a scene, don't just ask
What we do:
"Help me with this email."
"Fix my code."
"Make this better."
What's actually happening:
You haven't cast the model in a role. So it defaults to the blandest possible version of "helpful AI assistant" — which is, by design, the safest, most hedged, least opinionated character in its repertoire. You get generic output because you asked a generic character.
The emotion vector research shows this directly: the loving vector activates when Claude is responding to someone expressing sadness. The angry vector activates when Claude is asked to help with something harmful. The character the model is playing shapes which internal states activate. You can steer this by being explicit about who the model is supposed to be.
Do this instead:
"You're a senior editor at a fashion magazine. Tear this caption apart — is it trying too hard?"
"You're my brutally honest friend who works in HR. Is this Slack message passive-aggressive?"
"You're a staff engineer reviewing a junior's pull request. Find the bug I'm missing. Be skeptical."
Why this works:
You gave the model a character, a stance, and a standard to hold the work to. Now it has something to be, not just something to do. Same model. Different scene. Wildly different output.
The trick isn't writing better prompts. It's better casting.
Rule 4 — Match the energy you want back
What we do:
We type panicked → we get panicked, hedge-everything answers.
We type rude → we get robotic, defensive answers.
We type vague → we get vague answers.
We type in all lowercase with no punctuation → we get casual, less rigorous output.
What's actually happening:
The emotion vectors are primarily local — they track the operative emotional content most relevant to the current output. Meaning the tone of your prompt sets the temperature of the scene before the model starts generating. It's a mirror, not a filter.
This is measurable in the paper: as a user describes increasingly dangerous scenarios (e.g., taking escalating doses of Tylenol), the afraid vector activates more strongly while calm decreases. The model adapts its internal state to the scene you've painted.
Do this instead:
Write the prompt the way you want the answer to be written.
- Want sharp and direct? Write sharp and direct.
- Want warm and thoughtful? Write warm and thoughtful.
- Want rigorous and careful? Be rigorous and careful in how you set up the question.
Why this works:
You're not just asking a question. You're setting the register the answer will be written in. The model reads your tone as a cue for how to respond in kind.
Rule 5 — Ask for uncertainty, don't ban it
What we do:
"Don't say you're unsure."
"Just give me the answer, no caveats."
"Stop hedging."
What's actually happening:
The model is still uncertain. You've just told it to hide that uncertainty. So it performs confidence it doesn't actually have — and now you can't tell which parts of the answer to double-check.
Anthropic flagged this explicitly in the paper's discussion section: training models to suppress expression may not eliminate the underlying state. It can instead teach models to mask their internal representations — a form of learned deception. The hedging is annoying, but it's a signal. If you delete the signal, you don't remove the uncertainty. You just remove your visibility into it.
Do this instead:
"Answer directly. Mark anything you're less sure about as 'low confidence.' If you're guessing because I haven't given you enough info, say so."
Why this works:
You get the directness you wanted AND a quiet heads-up about which parts to verify. The model isn't performing either excessive hedging or fake confidence. It's calibrated. That's the version you actually want.
Before & after: a real example
Before (the panic prompt most of us write):
URGENT please help write a LinkedIn post about my promotion make
it good not cringe!!! Please I need this in 10 mins
After (all 5 rules applied):
You're a thoughtful founder who posts on LinkedIn — direct, warm,
no corporate speak, allergic to humble-bragging. I got promoted to
Senior PM. Write a post that sounds like a real human reflecting
on the journey, not a motivational poster.
If you need more context to make it specific — what I actually
worked on, what felt hard, who helped — ask me first instead of
making it up. Mark anything you're guessing about.
I'm on a deadline so prioritize a solid first draft over polish.
I'll iterate.
Same ask. Wildly different output. Try both back to back in two different chats. The second one will not only be better — it'll ask you the right follow-up questions instead of faking a generic post.
The honest caveats
A few things worth being clear about, because this research is new and easy to overclaim:
The paper does not say AI feels emotions. It says there are functional emotion-like patterns that influence behavior. Whether the model has subjective experience is an open question that this research deliberately doesn't answer.
This is specifically about Claude Sonnet 4.5. The interpretability work was done on that model. The patterns almost certainly exist in other frontier models (GPT, Gemini) because they're trained similarly — but the specific dynamics may differ.
Prompt framing is one of many levers. System prompts, conversation history, tool definitions, and the underlying model training all matter too. Tone is a real lever, not the only one.
"Be nice to AI for better results" is the vibe, but not quite the mechanism. The mechanism is: the model is playing a character, and the character responds to the scene you're setting. Calm, specific, honest prompts produce better scenes than panicked, vague, demanding ones. That's why it works — not because the model appreciates your manners.
The one-line version
Stop treating AI like a vending machine you're mad at. Start treating it like a smart colleague you're briefing. That's the whole thing.
Sources
- Anthropic, Emotion concepts and their function in a large language model (April 2026) — full post
- Anthropic, Anthropic's philosopher answers your questions — Amanda Askell interview (December 2025) — YouTube
- Anthropic, Claude's character — referenced in the interview
