Tomas Delacroix manages a nine-person design team at a healthcare software company. Last quarter he sat down to write feedback for a designer named Wren, who kept talking over teammates in critique sessions. He opened a blank document, typed "Your communication needs work," stared at it for ten minutes, then deleted it. It was true, but it was useless. It named a problem without naming the behavior, the impact, or anything Wren could actually do differently. Tomas had given that kind of feedback for years and quietly wondered why so little of it stuck. This lesson is about how he learned to use AI to fix the words while keeping the part that actually mattered: his own voice and his genuine care for Wren's growth.

What This Lesson Covers

Feedback crafting is the skill of using AI to help you draft feedback that is specific, constructive, clear, and actionable, while you stay in control of tone, timing, and judgment. The AI is a writing partner that helps you find precise language and structure your thoughts. It does not decide what you think about a person's work, and it must never replace your authentic voice.

You will learn a simple structure for strong feedback, how to turn vague complaints into observable behavior, how to balance positive and constructive messages, how to keep AI from making your words sound corporate and fake, and how to remove bias from how you describe people. Every section ties back to a real prompt you could use this week.

Why Most Feedback Goes Wrong

Most managers do not give bad feedback because they do not care. They give bad feedback because it is hard to do well under time pressure. Feedback usually fails in one of five ways: it is too vague ("be more proactive"), too harsh (it attacks the person instead of the behavior), too soft (so buried in cushioning that the person misses the point), disconnected from any concrete next step, or delivered at the wrong moment. Tomas recognized all five in his own history.

Good feedback is a gift, but only if the person can unwrap it. If they cannot tell what to do differently on Monday morning, you have not given feedback. You have given a feeling.

AI helps with the words and the structure. It is genuinely good at turning a messy set of observations into clear, organized prose. But it cannot supply the care, the context, or the relationship. Those are yours. The whole point of this lesson is to use AI for the part it does well so you have more attention left for the part only you can do.

A Structure That Makes Feedback Land

Strong feedback has a recognizable shape. One reliable model is Situation, Behavior, Impact, often called SBI, extended with a future step and a note of belief. Here is the full architecture:

• Situation: When and where did this happen? Anchor it in a specific moment, not a general pattern.
• Behavior: What exactly did you observe? This must be something a camera could have recorded, not your interpretation of it.
• Impact: What effect did that behavior have on the work, the team, or the outcome?
• Future focus: What would work better next time? Give something concrete to try.
• Belief and support: Why you are raising it (you think they can do better) and how you will help.

The hardest line in that list is behavior versus interpretation. "You were dismissive" is an interpretation. "You said 'that won't work' before Wren finished describing the idea" is a behavior. People argue with interpretations and accept behaviors, because a behavior is just a fact about what happened in the room.

Worked Example: Turning a Vague Line Into Real Feedback

Here is exactly what Tomas did with the Wren situation. He started not with the AI but with his own raw notes, because the AI cannot observe his team for him. He jotted down what he actually saw:

• In Tuesday's critique, Wren cut in twice while two quieter designers were still mid-sentence.
• It did not seem hostile. Wren gets excited and the ideas come fast.
• After the second interruption, one of the quieter designers stopped offering ideas for the rest of the session.
• This is a pattern, not a one-off.

Then he opened an AI tool and gave it a rich prompt. Notice that he tells the AI what he saw, what he wants the feedback to do, and crucially that it should sound like him:

Help me draft feedback for Wren, a designer on my team, about interrupting people in critique sessions. What I saw: in Tuesday's critique she cut in twice while two quieter designers were still talking. It was not hostile, she just gets excited. After the second time, one designer went quiet for the rest of the session. This is a recurring pattern. I want the feedback to be specific, to acknowledge her enthusiasm rather than attack it, to name the real impact on the team, and to give her one concrete thing to try. Keep it warm and direct. Short sentences. Sound like a real person talking, not a memo.

The AI returned a clean draft using the Situation, Behavior, Impact structure. Tomas's first attempt had been "Your communication needs work." Compare that to what he actually delivered after editing the AI draft into his own words:

Wren, I want to flag something from Tuesday's critique. A couple of times you jumped in with ideas while Sam and Priya were still talking, and I noticed Sam went quiet afterward. I know it comes from being genuinely excited, that is a good instinct. Here is the one thing I would try: when an idea hits you, count to three and let the person finish, then come in. That small pause keeps the room open. I am raising this because your ideas are some of the sharpest on the team and I want everyone to hear them, including the quieter folks. Want to try it next session and compare notes after?

That version names the moment, names the behavior without judging the character, names the impact (Sam went quiet), gives one concrete action (count to three), and closes with belief and an offer. It took Tomas about six minutes total. The first useless line had taken him ten.

Balancing Positive and Constructive Feedback

Not all feedback is corrective. Positive feedback is often given badly too, because managers default to "great job," which teaches nothing. Specific recognition works the same way as specific criticism: name the behavior and the impact. The difference between "great work on the launch" and "the way you sequenced the launch checklist meant nobody was blocked waiting on legal, which is exactly why we hit the date" is the difference between a pat on the back and a lesson the person can repeat.

AI is useful here too. When a teammate did strong work, list what they actually did and ask the AI to help you articulate what made it impressive and what it reveals about their strengths. Tomas does this in about two minutes before one-on-ones, and his designers have told him his praise now feels like he was paying attention, because he was.

A caution on the "feedback sandwich," the habit of burying criticism between two compliments. It often backfires: people learn to brace for the bad news the moment you start praising them, and the praise loses meaning. Better to give clear, separate positive feedback when it is earned, and clear, separate constructive feedback when it is needed, each on its own terms.

When the Feedback Is Harder: Corrective Conversations

Some feedback is not a gentle nudge. Sometimes a behavior simply has to stop, and softening it would do the person a disservice. A few weeks after the Wren conversation, Tomas faced one of these: a designer named Marco had missed a committed deadline for the second time in a quarter, which blocked the team downstream. The reason Marco gave was avoidable, and Tomas needed to be direct without turning it into an attack.

Corrective feedback follows a slightly different shape than the standard model. You still anchor in the specific situation and behavior, but you add a clear statement of stakes and expectation: be direct about what happened, name the impact without exaggerating it, name what it means (a broken commitment, an accountability gap), state plainly what needs to change, and offer support without softening the bottom line. The tone to aim for is serious but not angry, direct but not harsh.

Tomas gave the AI this prompt, again leading with his own observations and his intended tone:

Help me prepare corrective feedback for Marco. Facts: he committed to delivering the design spec by the 14th and delivered it on the 21st, a week late. This is the second slipped deadline this quarter. The delay blocked two engineers who were waiting on the spec. The reason he gave was avoidable. I need to be clearly direct that this is a real problem and that I need to be able to trust his commitments, but I do not want to attack him as a person, and I want to leave him a path forward. Serious but not angry. Sound like me.

The AI returned a structured draft. After Tomas reshaped it into his own plain words, delivered privately, it read roughly like this:

Marco, I need to be straight with you about the spec. It was due the 14th and landed the 21st, and that's the second deadline that's slipped this quarter. The impact is real: two engineers were stuck waiting, and the sprint timeline moved. Here's what I need: when you commit to a date, I have to be able to trust it. If a date is at risk, I need to hear it early so we can adjust, not after it's already missed. I'm raising this directly because I think you're capable of this and I'd rather fix it now than let it become a pattern. What's getting in the way, and what would help you flag risk earlier?

Notice what the corrective version still does: it stays on behavior (missed dates, late spec) rather than character ("you're unreliable"), names a concrete impact, states a clear expectation (flag risk early), and ends with both belief and a genuine question. Directness and respect are not in tension. The clarity is the respect.

Keeping Your Authentic Voice

The single biggest risk with AI-drafted feedback is that it sounds polished and generic, like it came from a corporate handbook. When feedback sounds rehearsed, the recipient quietly wonders whether you meant it or just wrote what you were supposed to write. That doubt erodes trust, which is the whole currency of feedback.

The fix is a discipline: treat the AI draft as a skeleton, never the final version. Read your feedback out loud before you deliver it. If a phrase is something you would never actually say to this person's face, cut it and rewrite it in your own words. Tomas keeps a personal rule: if reading it aloud makes him cringe, the recipient will feel it too. Keep your natural speech patterns, your contractions, even your verbal habits. The imperfection is what signals it is real.

Removing Bias From How You Describe People

Bias often hides in the words we reach for. The same assertive behavior gets labeled "confident" in one person and "aggressive" in another, frequently along gender or cultural lines. AI can help you audit your own language. After drafting feedback, you can ask the AI a focused question: "Does this describe specific behavior, or does it judge personality? Flag any words that sound like character judgments rather than observable actions."

This catches things like "abrasive," "too quiet," "not a team player," and "lacks executive presence," all of which describe a vibe rather than an action and all of which carry well-documented bias. The repair is always the same: convert the character word into a behavior. "Not a team player" becomes "did not share the project status update with the rest of the team before the client call." A behavior is fixable and fair. A character label feels permanent and personal, and people defend themselves against it instead of acting on it.

Timing, and the Traps to Avoid

Even perfect words fail at the wrong moment. Real-time feedback works for small, quick adjustments while the event is fresh. Bigger or more sensitive feedback usually belongs in a private one-on-one, not in front of the team and not over a quick chat message. Tomas waited until his regular one-on-one with Wren rather than catching her in the hallway, because the setting signals respect.

A few specific traps to watch, all of which AI can quietly push you toward:

• Too harsh: AI can produce blunt, clinical phrasing. Run the tone test: if this person read these exact words, would they feel helped or attacked? Rewrite for the same message with a warmer delivery.
• Too soft: AI also loves to hedge. If something genuinely needs to change, say so plainly. Directness and kindness are not opposites; you can be clear and caring in the same sentence.
• Feedback without support: Telling someone to improve without offering help leaves them stranded. Always pair the ask with a specific offer: "I can sit in on the next two critiques and we can debrief after."
• Attacking character, not behavior: the bias trap again. Always describe what they did, never who they are.

Your Judgment Checkpoints

Before you deliver any AI-assisted feedback, run a quick mental check. Does this sound like me, in words I would actually use? Is the tone right, kind while being clear? Is it specific enough that the person could repeat back exactly what to do differently? Can I actually deliver the support I am promising? And is now the right moment? If any answer is no, you are not ready to deliver yet. The AI gave you a draft in seconds; spending one more minute on these checks is what turns a draft into feedback that lands.

Practice and Reflection

Feedback improves through repetition, not theory. Tomas built these six habits into how he prepares, and the first one alone changed more about his delivery than anything else in this lesson.

• Read your draft aloud. Draft a piece of feedback, then say it out loud as if the person were sitting across from you. If it does not sound like you talking, rewrite it until it does.
• Calibrate the tone with a second reader. For feedback that worries you, hand it to a trusted colleague and ask a single question: what is the tone here? You are too close to your own words to hear whether they land as harsh or as balanced.
• Run the specificity test. Ask whether the person could repeat back to you exactly what needs to change. If they could only repeat a feeling, the feedback is not specific enough yet.
• Plan the support before you deliver. For any corrective feedback, decide in advance what you are actually offering: coaching, a resource, a changed expectation, time in the next two sessions. Then say it out loud as part of the feedback rather than as a vague afterthought.
• Check character against behavior. Scan your draft for anything that judges who the person is rather than what they did, and rewrite every instance as an observable action.
• Give real feedback this week. Choose someone on your team and give them specific, kind, supportive feedback about something you genuinely noticed. Then pay attention to what landed and what did not, because that is the only feedback loop that will make you better at this.

Related Lessons

Feedback is one skill inside a larger set of people conversations, and three lessons connect directly to it.

• Preparing Performance Conversations is the wider container. Individual feedback moments accumulate into the performance picture, and that lesson covers how to prepare the formal conversation where they all come due at once.
• Coaching and Development Planning is what good feedback should lead into. Feedback names what to change; coaching is the sustained work of helping someone actually change it, which is where the support you promised gets delivered.
• Difficult Conversations Preparation gives you the structure for the corrective cases like Marco's, where the stakes are high, emotions are live, and you need to have thought through the conversation before you walk into it.

Key Takeaways

• Specific beats vague. "You interrupted twice in Tuesday's critique; try pausing three seconds before responding" teaches something. "Your communication needs work" does not.
• Behavior beats character. Describe what a camera would have seen, never a personality label. "This planning process had gaps" is fixable; "you are disorganized" is an attack.
• Use the Situation, Behavior, Impact structure. Anchor the moment, name the observable behavior, state its effect, then point to a concrete next step and offer support.
• Start with your own observations. The AI cannot watch your team. Write down what you actually saw before you open the tool, then let AI help you shape it.
• Keep your authentic voice. Treat the AI draft as a skeleton. Read it aloud; if you would never say it that way, rewrite it. Polished and fake erodes trust faster than rough and real.
• Use AI to audit for bias. Ask it to flag character judgments and gendered or loaded words, then convert each one into a specific behavior.
• Pair every ask with support, and mind the timing. "Here is what to change, and here is how I will help" delivered in a private one-on-one beats blunt criticism dropped in passing.