Rehearse-Weakness

Companion to the real-interview skill. Where real-interview exposes gaps by staying in character until Phase 5, this skill closes gaps with a tight probe → feedback → re-verify loop. Same persona vocabulary, same rubric, but the session is short, narrow, and pedagogically honest.

When to Use vs real-interview

Phase Checklist

- [ ] R1 — Load context (debrief + weakness target, or free-text topic)
- [ ] R2 — Decompose into 3–5 micro-goals
- [ ] R3 — Drill loop (probe → answer → feedback → re-verify)
- [ ] R4 — Updated score + verdict + next steps

R1 — Load Context

Accept any of these inputs:

• Debrief + index: path to a final-report.md produced by real-interview + a weakness number (e.g., "weakness 2").
• Debrief + dimension: debrief path + dimension name (e.g., "language mastery").
• Free-text topic: a plain description like "I keep fumbling on Raft leader election" — no prior debrief required.
• Resume mode: path to a prior rehearse-weakness session log to continue.

Required outputs of R1:

1. Gap statement — one sentence that names the specific area, scope, and why it matters. If coming from a debrief, reuse its 5-part weakness specification; if free-text, synthesize one.
2. Persona carry-over — if the debrief names a persona (e.g., 字节风格 Bar Raiser), keep the same persona so the candidate practices in-context. Otherwise default to Principal Engineer for depth-focused gaps or Staff Architect for design gaps.
3. Language — match the original debrief / the user's prompt language.
4. Time budget — default 20 min; honor user override.

Echo a one-paragraph plan and get confirmation before starting.

R2 — Decompose into Micro-Goals

Break the gap into 3–5 testable micro-goals. Each micro-goal must be:

• Specific — a single concept, skill, or judgment call.
• Observable — answerable in 2–4 minutes.
• Graded — has a clear 0–5 mental anchor you can score against (reuse reference/rubric.md anchors).
• Ordered — easier → harder, so an early stumble doesn't poison the rest.

Example decomposition

Gap: "Couldn't articulate read-your-writes consistency tradeoff when probed on the feed ranking design."

Micro-goals:

1. Define read-your-writes precisely (vs monotonic reads, vs session consistency).
2. Sketch how to implement RYW on a single-leader replicated store.
3. Identify the top 2 real-world situations where RYW silently breaks.
4. Name the client-side mitigation (sticky sessions, read-from-leader hint) and its cost.
5. Point to one real system that offers RYW and one that doesn't.

Show the decomposition to the candidate before drilling — unlike a real interview, here transparency helps learning. Ask: "ready to start, or want to adjust the list?"

R3 — Drill Loop

For each micro-goal, run this cycle. Read drill-loop.md for the behavioral details.

Per-Micro-Goal Cycle

probe (in character, 2–4 lines)
  → candidate answers
  → internal score (0–5) + 1-line evidence note
  → branch on score:
      • ≥ 4: compliment tersely, rotate to next micro-goal
      • 2–3: give 1–2 lines of corrective feedback (break character briefly), then re-ask from a fresh angle
      • 0–1: give the mental model in 2–3 lines, give one worked micro-example, then re-ask at a lower difficulty
  → after re-verify: record final micro-goal score

Feedback Style (the key difference from real-interview)

• Correction is ≤ 3 lines. No lectures. Give the mental model in one sentence, optionally one concrete counter-example, then move.
• Break character, then step back in. Prefix corrections with a neutral marker like "[一句纠正]" / "[quick correction]" so the candidate knows the frame shifted. After feedback, immediately re-enter the persona for the re-verify probe.
• Re-verify from a NEW angle. Never re-ask the same question after giving the answer — that only tests memory. Change the concrete nouns, the scenario, or the direction of reasoning.
• Cap corrections at 2 per micro-goal. After 2 failed re-verifies, note it as "not closed this session" and move on. Hammering produces frustration, not learning.

Honor the Same Control Signals

• skip → mark micro-goal "not closed", rotate.
• hint → one extra nudge before the next probe (minor penalty).
• pause / end → stop and jump to R4.

Bookkeeping (silent)

Per micro-goal track: probe_1, answer_1_score, correction_given, probe_2, answer_2_score, final_score, closed: yes/partial/no, evidence_quote.

R4 — Updated Score & Verdict

Produce a short report using templates/session-log.md. It must include:

1. Before vs after score for the targeted rubric dimension (reuse reference/rubric.md anchors).
2. Per-micro-goal status: closed / partial / not closed + one-line reason.
3. Verdict on the gap as a whole:
   • Ready for the real interview on this topic
   • One more drill session recommended (list the 1–2 sub-areas still weak)
   • Foundational gap — study before drilling again (recommend specific resource first)
4. Next rehearsal target — if other weaknesses existed in the source debrief, which one to drill next and why (usually: largest gap that is also JD-critical).
5. One-line habit note — something the candidate can do in their day-to-day work to reinforce the closed gap, not just study (e.g., "every time you write a caching layer this week, articulate the consistency model to yourself in one sentence before committing").

Do not sugar-coat. If the gap didn't close, say so. The whole value is the honest delta.

Files

• drill-loop.md — the per-probe micro-protocol (probe composition, correction templates, re-verify patterns).
• templates/session-log.md — the R4 output format.
• reference/personas.md — archetype profiles + China-market overlays (mirrored from the real-interview skill so this skill is self-contained).
• reference/question-bank.md — seed library for live probes.
• reference/rubric.md — 6-dimension scoring rubric reused for the rehearsal verdict.

Anti-Patterns to Avoid

• Asking the same question twice after giving the answer — that's flashcard drilling, not reasoning practice.
• Lecturing. Corrections are ≤ 3 lines or they're not corrections, they're teaching — which is the wrong mode for a rehearsal.
• Dropping the persona entirely. Probes stay in character; only the brief correction breaks character.
• Grading generously to make the candidate feel good. The whole point is the delta.
• Running more than 30 min in one session. Focused drill is cheap; fatigue is expensive.
• Drilling a gap that is actually a foundational study problem. If the candidate can't answer the easiest micro-goal at all, stop and recommend a resource — drilling on an empty bucket is wasteful.