feat(prompts): tune Miss Pepper review voice [DEV-249]

[brodkin]

Summary

Reframes Pepper/Miss Pepper from two personas swapped by outcome to one voice with range — kills the model's tendency to self-narrate the persona switch instead of inhabit a single voice.

Resets the cliff: bookends still run loose at full register, but the middle is now plain-spoken Pepper rather than voice-free engineer prose — same person, lower volume, drag tokens off so the teaching can land. Adds a why-it-matters beat to the middle so the dev walks away with a principle, not just a fix.

Integrates the Miss Pepper character profile: principal-engineer-reviewing-juniors authority anchor, ruthless-to-code-not-the-developer ethical line, five technical-taste anchors, vocabulary palette of five tokens with usage functions, surprise-and-delight directive (mine the diff for setups it already offers), and an "embody, don't narrate" rule sharpened under the one-voice frame.

Voice/character layer only — no changes to review logic, intent verification, auto-fail rules, or test review behavior.

Closes DEV-249

Validation

Three Sonnet 4.5 simulation runs against a fixed test scenario (fictional Go event-ingestion PR with three real issues planted in the diff).

Beat	Iter 1 (pre)	Iter 2 (post)	Iter 3 (post)
Identifies all 3 real issues	✅	✅	✅
Cliff structure holds	✅	✅	✅
Plays at least one comic setup the diff offers	❌	✅ (X-Pepper-Deduped header)	✅ (go test -race teaching beat)
Screenshot-worthy specific praise	partial	✅	✅
Quotable teaching beat in middle	❌	✅	✅

Iter 2 sample of the surprise-and-delight beat playing:

Fourteen headers agree with me. One header borrowed my name. I'm flattered, truly, but it needs to be X-Spice-Deduped before this touches production.

Smoke test

Per Pepper deployment convention: a follow-up minor PR will exercise this prompt under the new v1 to verify real-world performance.

Test plan

• Pepper reviews this PR and either approves or routes to comment-and-assign — the prompt change should not block its own merge.
• Verified against DEV-249.
• After merge: v1 tag moves to this commit; smoke-test PR opened against the new v1.

[pepper-pr-review]

Verified against DEV-249 (Linear) — aligned. This prompt change delivers exactly what the issue calls for: reframes Pepper/Miss Pepper from two-persona swap to one voice with range, sharpens the cliff structure (loose bookends around a plain-spoken teaching middle), and integrates the character profile (ethical line, technical-taste anchors, vocabulary palette, surprise-and-delight directive). The "embody, don't narrate" rule at line 60 is the move that kills the self-narration tendency the validation runs surfaced — Miss Pepper stops announcing the bit and just is the bit.

No automated evals, but the three manual simulation runs documented in the PR body show measurable improvement (surprise-and-delight beats landing, quotable teaching moments, screenshot-worthy praise) without sacrificing technical rigor. Since this is voice tuning with no changes to review logic, auto-fail rules, or verification paths, manual validation is appropriate here. Building eval infrastructure would be valuable for future prompt changes, but it's not blocking this one.

— Pepper (rereading "The drag tokens drop in the middle because they'd compete with the teaching" like it's the line that finally makes the whole thing make sense)

When you're ready for another look, comment @pepper review.