Quick Answer
AI drafts difficult messages badly for a specific reason: it is tuned to be agreeable, and an agreeable draft of a hard message is a message that fails. The no gets buried, the apology turns into an excuse, the feedback softens into nothing. These 18 new prompts fix that by forbidding it directly, with instructions like the no must be clear by the second sentence, and play the difficult version of this person rather than a reasonable one. The library now holds 692 prompts.
There is a message sitting in your drafts folder. You have rewritten it four times. It declines something, or chases something, or tells someone that the work is not good enough, and every version reads as either too soft to work or too blunt to send.
Eighteen prompts went into the library this week for exactly that message. They cover saying no, renegotiating scope, chasing a late invoice, replying to an angry email, apologising for something that was your fault, telling your manager something is broken, ending a client relationship, and reviewing a colleague's code without starting a fight.
They share one design decision, and it is the reason they work. Every one of them tells the model to stop being nice.
What Are the 18 New Prompts For?
They spread across eight categories and group into four jobs.
Declining and renegotiating. Say No Without Damaging the Relationship, Decline a Meeting and Offer Something Better, Renegotiate Scope After the Brief Has Changed, End a Client Relationship Without a Fight, Say No Across a Hierarchy Gap in Southeast Asia.
Conflict you did not choose. Draft the Reply to an Angry Email Before You Have Calmed Down, Reply to a Public Complaint Without Escalating It, Write an Apology That Does Not Make Excuses, Chase a Late Payment Without Damaging the Relationship.
Saying it upward. Build the Case for a Raise Before You Ask, Raise a Problem With Your Manager Without Sounding Like a Complaint, Give Upward Feedback Without Risking the Relationship, Prepare for the Objection You Are Most Afraid Of.
Before and after the conversation. Decide Whether the Conversation Is Worth Having, Rehearse a Hard Conversation Before You Have It, Find the Sentence That Is Doing the Damage, Write Code Review Feedback People Act On, Negotiate Payment Terms With a Regional Supplier.
All eighteen use labelled variables in the usual way, so you supply your own situation rather than adapting somebody else's example.
Why Does AI Draft Difficult Messages Badly?
Because being agreeable is the model's default setting, and a difficult message is precisely the case where agreeableness is the wrong answer.
Ask a chatbot to help you turn down a request and watch what comes back. There is usually a warm opening paragraph, an expression of how much you value the relationship, a sentence that technically contains the refusal, an apology, a second apology, and an offer to help in future. The recipient can read the whole thing, feel good about it, and reply asking when you can start. Nothing in it is wrong. It simply does not do the job.
This is not a bug you can prompt your way around with please be more direct, because the pull toward pleasantness reasserts itself two paragraphs later. What works is naming the specific failure and banning it. Every prompt in this batch carries at least one line like that: the no must be clear by the second sentence, do not stack apologies, no sentence whose purpose is to make me feel better, do not open with praise I do not mean.
There is also a reason to explain the stakes rather than just issue the rule. Anthropic's prompting best practices note that giving context or motivation behind your instructions, including why the behaviour matters, helps the model understand your goal and produce more targeted output. That is why these prompts say what the constraint is for. Telling a model not to bury the refusal works better when it also knows that a buried refusal gets read as a yes.
The same instinct runs through the rehearsal prompt. Rehearse a Hard Conversation Before You Have It explicitly instructs the model to play the other person as you described them, defensive and long-winded, rather than as a reasonable version. Left to itself a model will roleplay a counterpart who concedes at the second exchange. That is a comfortable rehearsal and useless preparation.
How Do You Say No Without Damaging the Relationship?
You put the no where they will find it, and you stop apologising for it.
Say No Without Damaging the Relationship returns three versions at different firmness levels rather than one, because the right level depends on context you have not explained and probably cannot. Alongside each version it tells you what that phrasing signals about your availability in future, which is the part people miss. A very soft no to a recurring request is not a kindness. It is an invitation to ask again next month.
Two of its rules do most of the work. It forbids stacking apologies, allowing one acknowledgement at most. And it refuses to invent a reason you did not give it, or to dress your real reason up as something more impressive. If your reason is one you should not share, it says so and gives you a version that declines without a reason, rather than a fabricated one you then have to remember.
Decline a Meeting and Offer Something Better applies the same logic to the calendar, with one addition worth noting. It gives you a version for leaving a recurring meeting permanently rather than skipping it once, which is the version almost nobody writes. Skipping costs a message every week. Leaving properly costs one message ever.
Renegotiate Scope After the Brief Has Changed is the commercial version of the same conversation, and it opens by building a factual two column comparison of what was agreed against what is now being asked for, with no adjectives at all. The gap makes the argument. It also bans any sentence implying bad faith, because scope creep is usually accumulated small requests rather than a scheme, and treating it as a scheme turns a solvable conversation into a dispute.
Which Prompts Help When You Are Angry?
Three, and one of them does something deliberately strange.
Draft the Reply to an Angry Email Before You Have Calmed Down asks for two outputs. The first is the reply you want to send, written properly, with the corrections and the receipts. You are not going to send it. The point is to get it out of your head and onto a screen, because a reply you are composing mentally at eleven at night will keep being composed until it exists somewhere. Seeing it written down usually removes the urge.
The second output is the one to send: under 150 words, opening by agreeing with whatever they are right about without qualifying it, correcting the factual error exactly once in neutral language, and proposing the next step. It then tells you which sentence is most likely to reignite the exchange, and whether this should be a call rather than an email.
Write an Apology That Does Not Make Excuses enforces an order and bans a list. No sentence may begin with an explanation of why it happened. No passive voice about your own actions, so no mistakes were made. No conditional apology, so no if this caused any inconvenience. No promise that it will never happen again, which nobody believes. Its last instruction is the one that earns the prompt its place: it asks the model to tell you honestly whether anything in your description is self defence wearing the grammar of contrition. That is very hard to catch in your own draft.
Reply to a Public Complaint Without Escalating It starts from a reframe. The audience is not the complainer, it is the hundred people reading who will never comment. Its fourth section asks what must stay out of the public reply even though it is true and in your favour, which is where most brands lose. Being publicly right about a customer's mistake costs more than the refund would have.
How Do You Raise Something With Someone More Senior?
Carefully, and sometimes not at all.
Give Upward Feedback Without Risking the Relationship is the only prompt in this batch that may tell you to do nothing. Before drafting anything it weighs what changes if you speak, what happens if you stay quiet, and whether your standing and your workplace can carry the conversation. If the honest answer is not yet, or not from you, it says so and tells you what would need to be different. A prompt that always encourages you to speak up is not being useful, it is being encouraging.
That caution is not timidity, it is about who can afford to speak. Amy Edmondson's study of 51 work teams, published as Psychological Safety and Learning Behavior in Work Teams in Administrative Science Quarterly in 1999, found that team psychological safety, the shared belief that the team is safe for interpersonal risk taking, predicted learning behaviour, while team efficacy did not. Speaking up is a function of whether the environment makes it survivable, not of how confident the individual is. Reading your environment before you speak is the right first step, and it is the step most advice skips.
Where the conversation is worth having, Raise a Problem With Your Manager Without Sounding Like a Complaint does something harder than it sounds. It states the situation in three sentences of pure fact with no evaluative words at all: no overwhelming, no unsustainable, no unfair. Strip the adjectives out and the facts usually make the case better than the adjectives did. It then names the cost to the business before the cost to you, because both are real and only one changes decisions.
Build the Case for a Raise Before You Ask turns each achievement into a fixed shape, what changed, what it was before, what it is now, who benefited, and tells you to go and find the number where you have not supplied one. Its second step is the actual argument: grouping your work into the job you are doing now as opposed to the job you were hired for. That gap is what raises are granted for. Effort is not.
If you want to test how different models handle this kind of drafting before committing to one, the free tier comparisons at WhichAIBest cover current limits, and SpotFreeAI lists which tools have genuinely free plans rather than trials.
What Changes When the Conversation Crosses a Hierarchy Gap?
Almost all of it, and most prompt advice ignores this entirely.
Guidance about difficult conversations is overwhelmingly written from flat-office assumptions, where a direct no upward is admired as candour. That assumption does not travel. Hofstede's Power Distance Index puts Malaysia at 104, the Philippines at 94, Indonesia at 78 and Singapore at 74, against 40 for the United States and 35 for the United Kingdom. The index measures how far less powerful members of an organisation accept and expect that power is distributed unequally. A message that reads as refreshingly direct in one setting reads as a challenge in another.
Say No Across a Hierarchy Gap in Southeast Asia is built for the middle ground, because a flat no is costly and silence gets read as agreement. It returns three phrasings that decline clearly while leaving the other person's standing intact, and for each one it states what a senior local reader will actually understand from it.
Its second instruction is the important one. It flags which of your options risks being so indirect that it will be heard as a yes. That is the expensive failure. The work still gets assigned, and now you have apparently agreed to it. It also gives you the firmer version for when they ask a second time, tells you whether this belongs in writing or in person in your specific setting, and provides a sentence for a constrained yes with the constraint stated explicitly rather than implied.
Negotiate Payment Terms With a Regional Supplier follows the same discipline as our SEA pricing prompts: it refuses to give you exchange rates, standard terms by country, or typical margins, because those move and a model working from training data will be confidently out of date. It gives you the structure and you supply the current numbers. Its final instruction names the point at which you should stop pushing, because a supplier who agrees under pressure tends to recover the cost later as quality drift. Our prompts for Asian businesses guide covers the wider regional set.
How Should You Actually Use These This Week?
Pick one. Not eighteen.
Almost everyone reading this has one conversation they keep not having. Open Decide Whether the Conversation Is Worth Having and run that one through it. It is designed to force a decision rather than produce another balanced summary, because you already have the balanced summary and it is why you are stuck. It commits to one of three answers: have the conversation, let it go properly, or change your own behaviour. If it tells you to let it go, it also tells you what letting go involves, since deciding to drop something while remaining annoyed by it is not letting go.
One habit is worth adopting alongside whichever prompt you pick, and it costs nothing. Before you send any difficult message, read only the first two sentences and ask what a busy reader would conclude from them alone. Most people read no further before deciding how they feel. If your point is not in there, it is not in the message.
The library now holds 692 prompts across ten categories, all free and all usable without an account.
What Else Do People Ask?
Why is AI bad at drafting difficult messages by default?
Because it is tuned to be agreeable, and a difficult message needs the opposite. Ask a chatbot to help you decline something and you usually get warmth stacked on warmth, three apologies, and a no buried so deep the reader can reply cheerfully without having understood that you said it. The model is not confused about what you asked for. It is optimising for a pleasant exchange, which is the wrong target when the whole point of the message is that somebody will not enjoy receiving it. Every prompt in this batch works by naming that failure and forbidding it explicitly.
Should you let AI write a message to a real colleague?
Use it to draft and to pressure test, then send something you would recognise as yours. These prompts are built for that split. Find the Sentence That Is Doing the Damage deliberately refuses to rewrite your whole draft, because a full rewrite removes your voice from a message where your voice is the thing carrying the relationship. Rehearse a Hard Conversation Before You Have It never produces a message at all. The useful output is usually a list of what will go wrong and one prepared sentence, not a finished email.
What is the fastest way to check whether a difficult message will land badly?
Ask what a reader in their position will hear, rather than asking whether the message is clear. Clarity is not the usual failure. The usual failure is a sentence that reads as neutral to you and as an accusation to someone who is already defensive. Find the Sentence That Is Doing the Damage quotes back the three or four riskiest sentences, says what each one sounds like from the other side, and replaces only those. It also checks whether you actually asked for what you need or merely hinted at it and hoped, which is the more common problem.
Do these prompts work differently for Southeast Asian workplaces?
One of them is built for it specifically. Hofstede's Power Distance Index puts Malaysia at 104, the Philippines at 94, Indonesia at 78 and Singapore at 74, against 40 for the United States and 35 for the United Kingdom, so advice written for a flat office transfers badly. Say No Across a Hierarchy Gap in Southeast Asia gives you three phrasings, says what each signals to a senior local reader, and flags which one risks being so indirect that it gets heard as a yes. That last failure is the expensive one, because the work still gets assigned.
How many of these prompts do you actually need?
One or two. Almost everyone has a single conversation they keep not having, and one message type they redraft every month. Pick the prompt that matches each, save them where you can reach them in five seconds, and leave the rest. A prompt library is worth having because the right one is there on the afternoon a client sends an angry email, not because you worked through all 692 of them.
Sources: Amy C. Edmondson, "Psychological Safety and Learning Behavior in Work Teams", Administrative Science Quarterly, volume 44 number 2, June 1999, DOI 10.2307/2666999, for the study of 51 work teams in a manufacturing company and the finding that team psychological safety is associated with learning behaviour while team efficacy is not when controlling for psychological safety. Geert Hofstede's Power Distance Index scores as published by Clearly Cultural, for Malaysia 104, Philippines 94, Indonesia 78, Singapore 74, United States 40 and United Kingdom 35, and for the definition of power distance as the extent to which less powerful members of organisations accept and expect unequal distribution of power. Anthropic, Prompting best practices documentation, for the guidance that supplying context or motivation behind an instruction helps a model understand the goal and produce more targeted responses. All consulted 6 October 2026.