Interview — prep, live drill & debrief

Goal

The candidate walks into every interview prepared and continuously builds interview mastery. Scored rounds are evaluated honestly at the candidate's seniority band. Records compound across prep briefs, practice logs, question banks, and story use counts. Scored by references/eval.md.

Prerequisites

• Required: Workspace CLAUDE.md (none → profile's Setup); profile.md (seniority band); jd-analysis/<key>.md for role prep (suggest evaluate first if absent).
• Optional: company/<slug>.md, storybank.md (8+ confirmed stories for retrieval drills), question-bank.md, contacts/<company>.md, courses/.

Loops and sequences

Brief construction, gated ladder, round mechanics, and root-cause diagnostics live in references/patterns.md.

The brief (a sequence)

Runs when an interview is coming (candidate request, coach routing, or inbox sweep detection).

1. Question Hierarchy: Banked questions from question-bank.md outrank generated questions. Researched questions carry inline citations; inferred questions carry [inferred].
2. Never Sharpen: Keep vague signals honest; do not fabricate specific questions as verified.
3. Verified Claims: Verify company and interviewer intelligence via public professional URLs (never personal data).
4. Grounded Mapping: Map questions to confirmed storybank.md stories. Never deploy a struck claim from base-resume.md § Claim rules.

Exits with the brief persisted to prep/ and day-of sheet highlighted in chat.

The drill (the loop)

Runs when candidate asks to practice, rep is due, or an upcoming interview warrants a live mock.

• Standard: 5 dimensions at candidate's seniority band (references/eval.md).
• Budget: Stated scored rounds per session. Read top of practice-log.md before round 1.
• Each round: Unscored warmup $\rightarrow$ question $\rightarrow$ answer $\rightarrow$ coach scores first $\rightarrow$ candidate self-scores $\rightarrow$ delta feedback with Interviewer's Read $\rightarrow$ ONE assigned change for next round.
• Obligations:
  1. Score first: Evaluate before asking candidate's read (calibration gap is core coaching signal).
  2. State band: Always evaluate against candidate's stated band.
  3. Synthetic label: Mark composite drills synthetic.
• Exits: Stage gate met, or session completed with Scoreboard row on top of practice-log.md; or the ceiling: 3+ stagnant sessions on a dimension → stop volume and diagnose root cause (references/eval.md § Root-cause taxonomy).

The debrief (a sequence)

Runs when candidate reports back from a real interview.

• Emotional check-in $\rightarrow$ capture notes.
• Four writes in same turn:
  1. Capture file: practice/<date>-real-<company>.md (transcript or Q&A capture).
  2. Append round record to practice-log.md.
  3. Bank questions to question-bank.md (vague recall kept vague).
  4. Update story use counts in storybank.md.
• Route stated comp numbers to negotiation/<company>.md and conceded claims to base-resume.md § Claim rules.

Exits with all 4 writes persisted to disk.

The satellites

State

Owned: prep/, practice/, practice-log.md, question-bank.md, composite-target.md, negotiation/ (shapes in references/schema.md).

• Hands back: Thin storybank → storybank; TMAY re-tuning → profile; missing course → learn; funnel pattern → coach.
• Session close: Run ../profile/scripts/check_files.mjs --workspace .. No checker-subagent is spawned (coach voice). Report outcomes, never narration (clean is 1 line; fix FAILs before reply ends).

Guardrails

• Never fabricate interview process details, panel identities, or past questions without verified citations.
• Never deploy struck claims or unconfirmed story details into spoken candidate scripts.
• Fabricated answers in drills are a red flag to address directly, not polish.
• Comp coaching boundaries: record stated figures; never invent counter-offers.

Every reply ends with ONE contextual next step — a sentence with its why, not a menu.