The AI Nonviolent Communication Coach skill for Claude
You paste in the email you have written but do not dare send, or you describe the conversation you are dreading. It splits your message into observation, feeling, need and request, strips out the accusations and the disguised demands, and delivers a version that stays honest without making the other person defensive. It does not make the message softer or vaguer. It takes out the blame and leaves the content standing.
the lines above come from the zip on this page · SKILL.md is 16,911 bytes
Angry email in, message out
reg. G.001This is the worked example that appears in the SKILL.md itself, shortened here. Watch what happens to the draft: the skill does not rewrite it straight away, it first marks up what each sentence actually says. Only once you see which sentence is the problem does the rewrite become something other than a friendlier wrapping.
Context: my colleague always delivers his part of the monthly report at the last moment, which means I have to keep working into the evening every month. I have said something about it twice, and then it is better for one month. We are equal colleagues.
What the AI Nonviolent Communication Coach is
The AI Nonviolent Communication Coach is a free skill from our own library, free and with no account. A skill is an instruction file, SKILL.md, that gives an AI assistant a fixed way of working for one task. No software to install, no subscription, no link to your mailbox: one 2,740-word text file that tells Claude how a difficult message ought to be put together, what it needs to ask about, which words need to come out and what it must never fill in itself under any circumstances. What a SKILL.md actually is and why a text file steers an AI's behaviour is explained in what are Claude skills NL.
The problem this skill solves almost always starts at the same moment: there is a draft on your screen and you do not send it. You know it is too sharp, but you also know the friendly version achieves nothing, because you have already tried that twice. So it gets put off, and the subject grows. What happens then: the sentences that provoke defensiveness come out, and sentences come back in their place that carry the same content. Not softer, but aimed differently: at what happened rather than at who the other person is.
It is useful for work situations in the broad sense: a colleague who is always late, a manager with whom you want to set a boundary, a client whose email raises your blood pressure, a supplier who stretches agreements. The frontmatter also explicitly names the difficult-news conversation and saying no without damaging the relationship. It is part of the skill library NL that we make available for free from the AI and automation service, with no account and no sales email afterwards.
One thing is stated literally in the file, and it is the most important sentence on this page: the skill does not pretend the method changes the other person's behaviour. It changes the chance that you get heard. That sounds thinner than what training courses on this subject promise, and it is exactly what you should expect from it.
Four confusions, four tests
reg. G.002Rosenberg's method stands or falls on four distinctions that the SKILL.md names separately. Anyone who mixes them up writes an accusation in tidy words. On the left is what usually gets written, on the right what the skill turns it into, and below that the test you can use to check it yourself without AI.
OBSERVATION EVALUATIONYou deliver your parts late. OBSERVATIONThe last three times I received the piece on the day of the deadline. could a camera have captured this? no always, no never, no interpretation
FEELING PSEUDO-FEELINGI feel ignored and unappreciated. FEELINGIt makes me tense and I am disappointed. is there a judgement about what the other person did hidden in the word? then it is an interpretation, not a feeling
NEED STRATEGYI need you to report on Mondays. NEEDI need predictability and overview in that week. can this be met in more than one way? a strategy can only be met in one way, and that turns it into a negotiation
REQUEST DEMANDI expect you to be on time next time. REQUESTWould you be willing to see if it can happen two working days earlier, and if not, would you let me know what is getting in the way? what do you do if the other person says no? punishment, blame or distance means it was a demand, however friendly the sentence sounded
Why difficult conversations derail
Anyone who watches a conversation derail usually points to the tone afterwards. It is almost always in a handful of sentences you no longer notice yourself. There is a separate review mode for existing texts, and its checkpoints read like a list of everything that goes wrong in practice.
Diagnosing the other person. You apparently have no idea what that means for me. That is precisely the sentence the other person reacts to, not the rest of your email. The moment you say something about what is going on in the other person's head, the conversation becomes about that and no longer about the report. Every diagnosis is marked separately and then taken out.
Always and never. Generalisations invite exactly one response: finding the exception. Say 'you always do this' and back comes the one time it was on time last month, and you have lost the conversation over a detail that did not matter. Every generalisation is therefore replaced by a count: over the last three months.
The pseudo-feeling. Ignored, rejected, attacked, manipulated, unappreciated. They sound like feelings and they are not: they are interpretations of what the other person did, phrased in the first person. That is exactly why they provoke the defensiveness you wanted to avoid. Real feelings are words like worried, tense, disappointed, tired, relieved or irritated, and the skill notes in the output which interpretation it removed.
The disguised demand. I expect you to be on time next time. In friendlier words: I would appreciate it if you took this into account from now on. Both are demands the moment no is not an acceptable answer. The difference between a request and a demand is not in the sentence but in what you do when the other person says no. That is the fourth test from FIG.02, and immediately the only one you can only check afterwards.
Too much feeling in text. A rule of its own from the file, and a surprising one: in email and chat, long feeling passages land more heavily than intended. That is why the skill has a short variant in which observation and request are spelled out in full and feeling and need come together in one sentence. For a spoken conversation the opposite applies.
In diff form, with lines from the skill's example: what goes out and what comes back in its place.
What is actually in the SKILL.md
A skill is only as good as its instructions, so we simply describe them here. The file opens with a frontmatter stating when Claude should pick up the skill. Not just for the obvious terms like nonviolent communication, NVC, Rosenberg or compassionate communication, but also for sighs like: this is wearing me out, every time I say something about this he gets angry, I do not dare say it, I have written an email but do not dare send it, we keep talking past each other. That same frontmatter states the MIT licence, the version, and the limit: not to be used for unsafe situations, intimidation or abuse.
After that comes the theory, and it is unusually extensive for a skill file. It explains the four steps, the three confusions that determine whether it works, and the fourth test about the difference between a request and a demand. Then eight fixed steps: first ask about the factual situation and separate it from what the user thinks about it, mark every judgement and every disguised demand, translate pseudo-feelings, look for the need underneath the feeling and test whether it is not a strategy, formulate a request that no may answer, deliver the message in two forms, prepare the user for the other person's reactions, and warn if the situation cannot be solved with a conversation formula.
Before anything is written, the skill runs through a mandatory eight-point checklist: what factually happened, with when and how often, who the other person is and what your relationship is, what you yourself feel in your own words, even if that is anger, what you concretely want to change, whether you have said something before and what happened then, the channel, whether there is a power difference or an ongoing procedure at play, and what the consequences are if the conversation achieves nothing. It only asks about what is missing.
It contains eleven writing rules, from no dashes in running text to no therapy-speak: write it the way someone would say it out loud, not the way it appears in a training handout. And there is a separate review mode: give it an existing email, chat message or conversation note, and you get a version marked up sentence by sentence, a pseudo-feeling check, a demand check, a balance check and a rewritten version. That balance check is striking, because it tests the other direction too: if the text is softened so much that the message disappears, the file treats that as just as big a mistake as an accusation. If you want to learn to build an instruction file like this yourself, writing a SKILL.md NL explains how that structure works.
The file closes with eight things the skill never does and a source list of six titles. Those two lists shape the character of this skill more than all the steps put together, so they get their own place further down this page.
Eight blocks, fixed order
reg. G.003Every output of the skill has the same eight blocks, in the same order. That order is not arbitrary: each block answers the question the previous one raises. The mono lines underneath explain why the block is there.
The theory the skill rests on
The SKILL.md names its sources explicitly, and that is deliberate: so you can look them up and decide for yourself whether you agree. There are six of them, and they explain why the structure is the way it is.
The method comes from Marshall Rosenberg. He described nonviolent communication, usually abbreviated to NVC, in the book Nonviolent Communication: A Language of Life, first published in 1999, with later editions in 2003 and 2015. In 1984 he founded the Center for Nonviolent Communication, which publishes lists of feelings and universal needs that you can use alongside this skill. His starting point: people naturally like to contribute to the wellbeing of others, provided they are not pushed onto the defensive. Not a soft assumption, but the entire reason the four steps are the way they are.
The background comes from Carl Rogers. Rosenberg worked in his tradition, and On Becoming a Person from 1961 is listed as a source in the file: on empathic listening and unconditional acceptance. That explains why the skill has a variant in which you apply the four steps to the other person and ask whether your guess is right, instead of starting from your own observation.
The conversation types come from Stone, Patton and Heen. Difficult Conversations from 1999 splits every hard conversation into the facts conversation, the feelings conversation and the identity conversation. This is referenced the moment the core of the conflict is not about what happened but about how someone sees themselves: as a good colleague, a reliable supplier or a good parent. A better formulation does not help there; you need a different conversation altogether.
The high-stakes angle comes from Patterson and colleagues. Crucial Conversations from 2002 is about conversations with big consequences and opposing opinions. Where Rosenberg supplies the language, that book supplies the context: the higher the stakes, the sooner people fall back on silence or attack.
And there is a workbook. The Nonviolent Communication Companion Workbook is in the source list, with exercises for the four steps. That fits the character of this skill: the steps are simple to understand and hard to stick to the moment things get tense.
For the organisational side there are related skills in the same library. If it is about giving feedback to someone on your team, the Radical Candor Coach NL is the more direct route in, with situation, behaviour and impact instead of observation, feeling, need and request. If it is about why someone gets defensive before you have even finished speaking, the answer is often status or autonomy, and that is the territory of the SCARF Model Coach NL.
Even without installing the skill you can improve your own email today with these principles: take out the diagnoses, replace always and never with a count, and check whether no is an acceptable answer to what you are asking.
What the skill refuses
reg. G.004The SKILL.md contains a list of eight things the skill never does, and that list matters at least as much as what it does do. A message about a conflict carries your name. In conversation those rules play out like this: every rule is a request you might make, with the response the skill gives according to its own instructions.
Installing in Claude Code, Claude.ai or Codex
The zip contains a single folder with the SKILL.md inside it. Installing is a matter of putting the file in the right place, and that place differs per environment. SKILL.md has been an open standard since December 2025, so the same skill also works in Codex, Cursor and Gemini CLI. So you are not downloading a Claude file but a working instruction that any modern AI assistant can read.
- Unzip it into
~/.claude/skills/(or.claude/skills/in your project). - Claude then recognises the skill automatically as soon as you start talking about a difficult conversation or an email.
- You can also call it directly, with
/ai-geweldloze-communicatie-coach.
- Go to Customize and then Skills.
- Upload the zip there as a skill.
- Or paste the contents of SKILL.md into the project instructions of a Project.
- Open
AGENTS.mdin your repo. - Paste the contents of SKILL.md into it, or put SKILL.md next to it as a separate file and refer to it from
AGENTS.md. - Codex reads that along with every session.
After that, using it is simple: paste in your draft email, or describe the conversation in your own words, including your irritation. The rawer the better, because that is exactly the material the skill pulls the judgements out of. It only asks about what is missing. If you get stuck on the installation itself, the full route is in installing Claude skills NL. If you want to learn to set up this kind of work with your team, that is the subject of our AI training NL.
When to use it, and when not to
It is at its strongest for a planned conversation where you have time and the relationship matters: that is where the four steps really come into their own. For email and chat it switches back to the short variant, with observation and request spelled out in full and feeling and need combined in one sentence, because long feeling passages land more heavily in text than intended.
If you are still in the emotion yourself, the order changes. If you are still angry, it does self-empathy first: the same four steps aimed at yourself, before anything is written for the other person. If the other person is the one who is angry, it does not start with your observation but with a guess at their feeling and need, and it asks whether that is right. In a formal context, such as an appraisal conversation or a file note, it keeps the four steps but notes feeling sparingly and puts the weight on observation and request.
There are also situations where you are better off leaving it alone, and the file is unusually explicit about that. For intimidation, threats, discrimination or violence, the skill does not use this method as a solution, but points towards a confidential adviser, a manager, a lawyer or support services. It also gives no advice on forgiving or on staying in a situation, because that is not a communication question. And it refuses to use the four steps as tidy packaging for a demand, which in practice is the most common way this method gets misused.
One more honest limit, and it is stated literally in the file: the method does not change the other person's behaviour, it changes the chance that you get heard. If someone consistently breaks agreements and is still doing it after three conversations, the subject is no longer the delivery but the division of work, and someone else needs to be involved. That is not the method failing, that is the moment the method is done. If the conversation is about a clashing interest rather than a need that is not being heard, the Negotiator skill is the better route in.
Run it yourself or have it run
reg. G.005This skill is the free do-it-yourself version of work we also deliver as a service. Nothing is held back and there is no catch, but be clear about what a skill is: it teaches your AI how to do something, while every new session starts empty. It is not the engine, and it is not the memory either. You prompt, you supply the context about this person and this conflict again every time, you check the result. Anyone who wants that differently has two follow-up steps: hand over the engine, or sort out the memory.
where you are now The skill: you are the engine You run the AI Nonviolent Communication Coach yourself in Claude, Codex or Cursor. Costs nothing, works today, and you keep it entirely in your own hands: no trial period, no locked-off parts. The limit is your own time: it only happens when you prompt.
only if you come back The AI employee: usually not the next step here None of our roles fit here, and we would rather say that than sell it to you. The three roles we set up ready-made are the Quote Employee (sorting incoming requests and preparing draft quotes), the Sales Employee (prospect research and outreach drafts) and the Reporting Employee (summaries and weekly and monthly reports from your own data). What this skill delivers, a message in observation, feeling, need and request, fits none of those three. What is possible is a custom-built role via Mansotti, the company of which TheSEO is the trading name, but only if there is a stream of customer messages underneath it that needs answering every day. If this work stays a single session a year for you, skip step 2: you do not outsource a conversation with a colleague, and we do not offer that either; what does help is an AI that remembers your agreements and your tone, and that is step 3. What an AI employee does do is on that page.
everything from one source Jarvis: all your AIs work from the same company knowledge The skill teaches the AI, the brain is where the memory lives. Want all your AIs to work from the same company knowledge: that is Jarvis, the organisation's brain. It connects ChatGPT, Claude, Codex and your people to the same projects, core knowledge and decisions, so your next AI session does not start from zero. Jarvis selects context, delegates work and stores only what has been approved: client knowledge stays isolated and every step leaves a checkable trail. This skill also gets better because of it, since you no longer have to explain the tone you use towards clients and what has already been agreed again every session. What that delivers in practice, from the plans to your first week, you can read at Jarvis itself.
What Jarvis actually delivers
reg. G.006Step 3 deserves more than one paragraph, because this is the difference between a clever chat and a system you can build on. Jarvis is the organisation's brain: it remembers what your AIs need to know, divides up the work and keeps track of what happened. With difficult conversations you notice that at one very specific moment: when you have to explain the same situation for the third time before your AI can say anything useful.
We have been running on this system ourselves for months. Every agent session, every task and every decision is logged in it and can be read back. A new session therefore does not start blank: it first retrieves the recorded decisions, the running projects and the latest changes, and carries on where the last one stopped. So we are not describing a promise but the way of working we operate in ourselves, every day.
See the four plans at jarvis/pricing NL. Through the waitlist NL you only pass on your preferred plan, without obligation. That does not yet create an account, an order or a payment obligation. We discuss bespoke business work first.
The skills around it
reg. G.007A difficult conversation never stands on its own: there is a relationship underneath it and behaviour afterwards that either changes or does not. These skills from the same library each take on a different part of that.
The same conversation, a different angle
the messageTwo alternatives for the same situation, with a different structure.
Radical Candor CoachSituation, behaviour and impact instead of observation, feeling, need and request. More direct, and built for feedback within a team.SKILL NL NegotiatorFor the moment it is not about a need that is not being heard, but about a clashing interest.SKILLWhy the other person shuts down
beneath the surfaceSometimes the resistance is not in your words but in what is unintentionally at stake.
SCARF Model CoachStatus, certainty, autonomy, relatedness and fairness: the five buttons that make people defensive.SKILL NL Johari WindowWhat others see in you that you do not, and how you shrink that blind spot before you go into a conversation.SKILLGetting behaviour to change
after the conversationThe method changes the chance that you get heard. These three are about what is needed afterwards.
Fogg Behaviour Model CoachWhy behaviour does not happen: motivation, ability and a prompt all have to line up.SKILL EAST Framework Behaviour CoachMake it easy, attractive, social and timely, instead of saying something about it one more time.SKILL Kotter Change CoachFor when it is not one person but the whole department that needs to start doing something different.SKILLLooking further
the contextWhere this skill comes from and what else there is.
The whole skill libraryAll 100 free skills in a row, sorted by subject.HUB NL AI and automationThe service behind it: from separate skills to working automation in your business.SRV What are Claude skillsThe basics: what a SKILL.md is and why a text file steers an AI's behaviour.DOC NL Knowledge baseArticles about SEO, AI and online visibility, searchable.DOCFrequently asked questions
What does the AI Nonviolent Communication Coach skill cost?
Nothing. The skill is free, released under the MIT licence, and you do not need to create an account or leave an email address. You download a 6.7 KB zip containing a folder and a single file, SKILL.md, and that is the complete skill. There is no paid version and no sales email follows.
Does this skill also work in Codex, Cursor or Gemini CLI?
Yes. SKILL.md has been an open standard since December 2025, so the same file also works in Codex, Cursor, Gemini CLI and other tools that follow the standard. In Codex you unzip it into .agents/skills/ in your project, or into ~/.agents/skills/ for all your projects; Codex has supported SKILL.md directly since the open standard of December 2025. Putting the contents of SKILL.md into your AGENTS.md still works too. The instructions are simply readable text, so any AI assistant that accepts instruction files can work with it.
Will this make my message far too soft?
That is precisely the mistake the file rules out first. The skill does not make the message softer or vaguer: it takes out the blame and leaves the content standing. There is even a separate balance check in review mode, with the rule that a text softened so much that the message disappears is just as big a mistake as an accusation.
What is a pseudo-feeling and why does the skill remove it?
Words like ignored, rejected, attacked, manipulated and unappreciated sound like feelings, but they are interpretations of what the other person did. That is exactly why they provoke the defensiveness you wanted to avoid. The skill translates them into real feelings such as worried, tense, disappointed or irritated, and notes which interpretation was removed.
What is the difference between a request and a demand?
Not the wording, but what you do when the other person says no. If you respond with punishment, blame or distance, it was a demand, however friendly the sentence sounded. The skill tests every request against that and rewrites it if the test is not passed, because using the four steps as packaging for a demand is on the file's list of things it never does.
Can the skill also review an email I have already written?
Yes. Give it an existing email, chat message or conversation note, and you get a version marked up sentence by sentence, a pseudo-feeling check, a demand check, a balance check and a rewritten version following the standard structure, with the spoken and the written form. That is the fastest way to see exactly where an angry email derails.
What situations is the skill not meant for?
For unsafe situations, intimidation and abuse. That is a hard limit stated in the file: in that case the skill points towards help instead of a conversation formula, and the route is a confidential adviser, a manager, a lawyer or support services. It also never advises forgiving someone or staying in a situation, because that is not a communication question.
You send the email yourself
A perfectly worded message sitting in your drafts changes nothing. What you get here takes the writing off your hands, not the sending and not the conversation. What it does do is lower the threshold: you know which sentence was the problem, you have a spoken and a written version, and you know what to say when the other person comes back with 'I'm busy too'.
If you want to build this more broadly into your business, there are two routes that cost something and deliver something. In the AI training NL your team learns when to put an AI between two people and when not to. And if you want your agreements and your tone to still be known to your AI next month, instead of explaining them again every session, that is Jarvis, from €9 per month including VAT.