Feedback rephraser

Source: points/performance-review-rules.md (reframing kit) and points/core-rules.md (say what you like + what you would like).

Why this exists

Blunt feedback feels honest to the giver and burns trust with the receiver. The receiver hears "you are bad at X" even when the giver meant "the work needs Y." The fix is mechanical: lead with what you LIKE, then what you WOULD LIKE, anchored to the WORK, phrased as a QUESTION where possible.

The principle generalizes across every channel — annual review, Slack DM, 1:1 talking point, peer review comment, HR script. Same rewrite, same rules.

What to detect

Catch any sentence that does one or more of:

• Addresses a personal trait, not a behavior in the work ("you don't smile enough," "you're intellectually intimidating," "you're too blunt")
• Tells the person to stop something without naming what to do instead ("stop interrupting," "stop being late")
• Uses you are / you're too / you don't as the load-bearing verb
• Speculates about the person's emotional state ("you're burning out," "you don't care")
• Mentions personal context the writer hasn't been invited to discuss ("because of your divorce," "since the baby was born")
• Phrases a complaint as a declaration when a question would land softer ("this isn't working")

The five-move reframe

Every rewrite uses some combination of these moves:

1. Open with what you LIKE — at least one named, specific thing.
2. Re-aim at the WORK — not their personality, not their emotional state. The deliverable, the meeting, the client interaction.
3. Convert "stop X" to "I'd love to see more Y" — frame the positive behavior, not the prohibition.
4. Phrase as a question — "can we talk about how to...", "is it time to...", "worth thinking about whether..."
5. Cut all psychoanalysis — never speculate about why. If you suspect a personal reason, that's a separate, private conversation.

Reframing table

The patterns that show up most often. Detect any of these constructions and rewrite them.

Output format

Two passes, same shape as style-tells.

Pass 1 — Inventory

Blunt phrasings found: N

For each occurrence, return:

Line N: "the original blunt sentence"
  Detected: trait-attack | stop-without-replacement | psychoanalysis | declared-not-asked | personal-context
  Reframe: "the rewritten sentence"
  Why this lands better: one sentence

Pass 2 — Clean draft

The full text with every blunt line rewritten. No commentary in this section, just the clean text.

If the text is a written review, also flag any line that introduces bad news the writer has not raised before. The rephrase doesn't fix an ambush — the writer needs to discuss it with the person first.

When NOT to soften

Some feedback should stay sharp. Do not reframe if any of these apply:

• The behavior is a fireable offense (harassment, theft, lying to customers) — sharp written record is the point.
• The user is documenting rather than coaching — legal / HR context where understatement creates risk.
• The receiver is the user themselves writing a self-critique — self-reframing into corporate-speak is often the user dodging their own assessment.
• The user explicitly says "I want this to be blunt" or "this is HR documentation."

When in doubt, ask the user which context this is in.

What this skill is not

• Not a politeness layer. The reframes are tighter, not softer. They cut filler, replace personality attacks with work-focused asks, and convert declarations into questions. The receiver gets more signal, not less.
• Not a yes-and machine. If the feedback is wrong on the merits, the reframe won't save it. Flag that and tell the user the underlying claim needs work first.
• Not for upward feedback in hierarchical cultures where the issue isn't phrasing but channel — those need a different surface (skip-level, anonymous survey, exit interview), not a rewritten sentence.

The litmus test

Read the reframe aloud. Would you say it to the person across a coffee table without flinching? If yes, ship. If it still reads like a manager who's annoyed, you stopped at "what I would like" without first naming "what I like." Loop back to move 1.