Draft Feedback That Lands with SBI Structure and Tone Calibration

Workflow area: Leadership Ops · Tier: Foundation · Hands-on time: ~20 minutes

Feedback is a factual claim wrapped in a relationship. The factual part — situation, behavior, impact — is where AI genuinely helps: it separates observation from judgment, catches the adjectives that will start an argument, and lets you hear the same message at three temperatures before you pick one. The relationship part it cannot touch. And the catastrophic failure mode is specific: a model asked to "write feedback about X" will happily invent plausible examples that never happened. You supply every fact; the model supplies only structure.

When to use — and when not to

• Use it for feedback you keep postponingthe drafting friction is usually the block, and twenty minutes of structuring beats three weeks of avoidance while the behavior compounds.
• Use it for written feedback that will be read without youreview inputs, peer feedback requests, anything where tone lives entirely in the words.
• Use it when you are angrythe tone-calibration pass is a cooling-off ritual; the version you would have sent at 6pm rarely survives it.
• Skip it for praise in the moment"that migration plan was excellent" needs no framework; instant beats structured.
• Skip it for anything heading toward HRdocumented underperformance, conduct issues, and anything that could appear in a termination file follow pb-22, with HR review and enterprise tooling.
• Skip the delivery step entirelythis playbook drafts; the conversation is yours, live, with room for their side to change your view.

The recipe

1. 1.

   Write down only what you observed. Before any AI, three raw inputs: what happened (dates, artifacts, who was present), what the person specifically did or said, what the consequence was. If you cannot name a specific instance, you do not have feedback yet — you have a vibe, and no framework fixes a vibe.

2. 2.

   Run the structuring pass.

   Copy

   Structure my observations into SBI feedback (Situation, Behavior,
   Impact). I will deliver it verbally to [ROLE — e.g. "a senior
   engineer who reports to me", no names].
   
   My observations — the ONLY facts you may use:
   [PASTE OBSERVED FACTS]
   
   Produce:
   1. SBI DRAFT — Situation: when and where, specific. Behavior:
      only observable actions, no motives, no character words.
      Impact: concrete consequence on people, work, or timeline.
   2. FLAGS — every phrase in my input OR your draft that is vague,
      judgment-laden, mind-reading, or disputable by a defensive
      listener ("always", "unprofessional", "didn't care"), each
      with a factual replacement.
   3. GAPS — where my observations are too thin to support the
      claim; say what a specific example would need to contain.
   
   Hard rule: do not invent, embellish, or "improve" any example.
   If a section lacks facts, say so.
   

3. 3.

   Generate the three tone variants. In the same chat:

   Copy

   Now give me three variants of the SBI draft for a live
   conversation: (a) DIRECT — plain, minimal cushioning, for
   someone who prefers straight talk; (b) WARM — leads with care
   and shared goals, same facts; (c) COACHING — ends in an open
   question that invites their read of the situation. Keep every
   fact identical across variants. Mark any sentence that differs
   in factual content, not just tone — there should be none.
   

   Good output keeps the spine constant:

   Copy

   DIRECT: "In Tuesday's design review, you dismissed Kavya's
   concern about the retry logic before she finished — 'we've been
   over this.' She stopped contributing, and the concern turned out
   to be the bug we shipped. I need review pushback heard in full."
   
   FLAGS (from earlier pass): "you always shut people down" —
   unsupported generalization; one observed instance. Replaced.
   

4. 4.

   Pick the variant that fits the person, then make it yours. Read it aloud once. Swap any word you would not naturally say — delivering model prose verbatim sounds like delivering model prose, and the person will hear it.

5. 5.

   Deliver live, then listen. SBI is your opening, not the conversation. Their account may add facts you did not have; feedback that cannot survive new information was a verdict, not feedback.

Make it yours

• New manager: run your first ten pieces of feedback through the FLAGS pass even when you skip the rest — learning to see your own judgment-words is the durable skill, and the tool is training wheels for it.
• Leading leaders: add a line to the prompt — "the recipient is a manager; frame impact at the level of their team's outcomes and the example they set" — because behavior feedback to a manager is really system feedback.
• Cross-cultural or distributed teams: ask for a fourth variant calibrated for high-context communication, and have a trusted local peer read it; tone calibration by model is a starting point, not cultural competence.

Quality check

• Every fact in the draft appears in your inputdiff them; anything new is invented, and delivering an invented example to a real person is indefensible.
• Zero character words survived"sloppy," "unprofessional," "doesn't care" are verdicts; behaviors are filmable.
• The impact is real, not theoreticalif you cannot name who or what was affected, the GAPS section was right and you need another example first.
• The three variants share identical factstone changes temperature, never content; a variant that softens the facts is a different (and dishonest) message.
• You would say it face to facethe read-aloud test catches prose you would never own in the room.