The AI Prompts That Make a Feedback Conversation Easier to Prepare

Ask AI to help you draft hard feedback and the easy failure mode is asking it to make the message softer. It will happily oblige, hedging the language, adding qualifiers, wrapping the actual point in enough cushioning that it's genuinely possible to walk away from the conversation not sure what was actually said. That's not kindness. It's feedback that fails to land, dressed up as tact.

The more useful question isn't "how do I say this more gently." It's "does this message show I care about this person, and is the actual point still unmistakably clear."

Why does AI-softened feedback get harder to actually understand?

Ask an AI to make a piece of hard feedback sound gentler and it will comply: more qualifiers, more acknowledgment of context, more room for the recipient to read it charitably. Read the result back and it's noticeably harder to identify the actual point being made. The message has gotten kinder-sounding and less useful at the same time.

This isn't a quirk of AI phrasing. It's the same failure mode unstructured feedback has always had: conflating kindness with vagueness, when the two are actually independent, and a feedback message can be warm and unmistakably clear at once, or cold and vague, in any combination, which most people never think to check separately before hitting send.

Is AI actually useful for preparing hard feedback, or just for softening it?

Unaided, it's genuinely useful for flagging specific phrases in a draft that read as either needlessly harsh or so hedged they'll likely be missed, and for generating alternate phrasings once you've decided on the actual point you need to make.

It's weak at knowing this specific person, how they've responded to feedback before, what history exists between you, whether directness lands as respect or as an attack coming from you specifically given everything that's happened between you two so far. It has no access to that relationship history unless you describe it in detail, and even then it's working from your account of it, not theirs.

What's worth asking AI before you draft this feedback?

These four run in Claude, ChatGPT, or Gemini as written.

1. The clarity check. "Read this feedback draft: [paste it]. If you were the recipient, what specifically would you think I'm asking you to change? If it's unclear, say so rather than guessing generously." Why it works: this catches feedback that's been softened into ambiguity, where the AI, standing in for the recipient, genuinely can't identify the actionable point.

2. The care-and-challenge audit. "Score this draft on two separate axes: does it show I care about this person specifically, and does it clearly challenge the behavior that needs to change? Where is it weak on either axis?" Why it works: scoring the two axes independently catches the specific failure mode of thinking you're being kind (high on care, apparently) while actually just avoiding the challenge entirely, or vice versa.

3. The roleplay rehearsal. "Play [describe the person, defensive, likely to get emotional, likely to argue back] hearing this feedback for the first time. React realistically, then tell me afterward what part of my delivery would have landed worst." Why it works: rehearsing against a plausible reaction, while the stakes are still zero, surfaces where your own delivery would land badly before you find out in the real, much higher-stakes conversation.

4. The reversal check. "Here's feedback I received recently: [paste it]. Using the same care-and-challenge framework, was this actually caring, actually direct, both, or neither?" Why it works: running the framework on feedback you received, not just feedback you're giving, helps you separate a legitimately hard truth you're resisting from an actually unfair or poorly delivered message, a distinction that's much harder to make in the moment than in hindsight.

What's missing even after you've prepared well?

This preparation will make today's conversation go noticeably better than an unprepared one. It won't give you a standing method for every feedback conversation ahead, the one next month with someone else, the one in six months once this person has already heard your first round of directness and needs something calibrated differently. One well-prepared message isn't the same as a repeatable discipline for calibrating care and challenge every time.

How does Kim Scott's framework actually change the way feedback gets drafted?

Kim Scott, documented in Radical Candor, built a two-axis model: care personally, genuinely invested in this person as a whole human, not just their output, and challenge directly, willing to say the hard, specific thing rather than talk around it. Feedback that's high on care but low on challenge is what she calls ruinous empathy, avoiding the hard truth out of what feels like kindness but actually leaves the person without what they need to improve. Feedback that's high on challenge but low on care is obnoxious aggression, technically accurate but delivered in a way that damages the relationship and gets heard as an attack rather than help.

Before, generic prompting: asked to help soften a message about a team member missing deadlines repeatedly, a generic AI response adds qualifiers and context, "I know things have been busy, and I really appreciate everything you do, but," which reads as an apology wrapped around a complaint rather than direct, actionable feedback.

After, Scott's framework applied: the framework separates care from softening entirely. It keeps a genuine statement of care, specific and true, "I want you to succeed here and I think you're capable of more than this pattern suggests," and pairs it with a direct, unhedged statement of the actual problem, "you've missed three deadlines this month, and I need to understand what's actually happening so we can fix it," rather than folding the challenge into vague, deniable language. Care and challenge sit side by side, explicit, rather than the challenge getting diluted by the care.

A generic "make this softer" request has no reason on its own to protect both sides of that pairing, care and challenge, at once. The framework refuses to trade one for the other, structurally, regardless of how the softening request was phrased.

How do you get this framework running in Claude, ChatGPT, or Gemini?

Claude takes it as a Skill: Settings → Skills → Add skill, upload the .zip. It auto-invokes once your question matches the skill's description, or gets forced with /kim-scott-framework at the start of a message. In Claude Code, the same package installs into ~/.claude/skills for every project, or a project's own .claude/skills folder for just that one.

ChatGPT has no Skills equivalent, so the plain .md becomes a Custom GPT's Instructions: Create a GPT → Configure → Instructions, pasted in directly. Skip the default Custom Instructions fields, they cap at 1,500 characters; a Custom GPT's Instructions field runs to roughly 8,000, enough for the full framework.

Gemini has no dedicated upload option, so the content goes into the system prompt if your interface exposes one, or opens the first message instead: "Here is a thinking framework to apply throughout this conversation. Read it carefully, then answer using this framework's approach," with the framework text pasted below it.

What's the next step from here?

If there's a specific conversation you've been postponing, Difficult Feedback is a $49 tool built to structure exactly that conversation before you have it. For the underlying discipline, Scott's documented method is one of several in the People & Culture category, alongside Amy Edmondson and Laszlo Bock, each available as an .md file for Claude, ChatGPT, or Gemini.