Interview

Interview — phased conversational context review + fill

🚨 MANDATORY: Voice Notification

Before running the workflow, send:

curl -s -X POST http://localhost:31337/notify \
  -H "Content-Type: application/json" \
  -d '{"message": "Starting the interview. Scanning phases first."}' \
  > /dev/null 2>&1 &

What this skill does

Runs a phased conversational interview across every PAI context file. Phase 1 (foundational TELOS) is the core — the DA always reviews it first, even if files look "complete," because foundational context is never actually done. Only after Phase 1 does the interview move to IDEAL_STATE dimensions, preferences, and identity.

The phases

Review vs. Fill mode

• Fill mode (completeness < 80%): walk through the scanner's prompts, write answers to the file's structured slots.
• Review mode (completeness ≥ 80%): read the file contents to the principal first, then ask targeted questions — "Anything outdated? Anything missing? Sharpen or refine any of these?"

The scanner marks each target's mode based on completeness. The DA respects that mode in the conversation.

Workflow

Step 1 — Scan

Run the scanner to see phase breakdown and current state:

bun ~/.claude/PAI/TOOLS/InterviewScan.ts

The scanner orders items phase-first (Phase 1 always before Phase 2). Present the per-phase summary to the principal:

"Your setup is 85% overall. Phase 1 foundational TELOS is at 100% — but every file is worth a review pass. Phase 2 IDEAL_STATE is at 59%. Phase 3 preferences mostly good except FOOD_PREFERENCES at 7%. Starting with Phase 1: MISSION. Ready?"

Step 2 — Walk Phase 1 first

Do not skip Phase 1. Even if every foundational TELOS file scores 100%, walk through each in priority order. These files are never truly "done" — a quarterly review pass is what keeps them current.

Phase 1 order (by leverage): MISSION → GOALS → PROBLEMS → STRATEGIES → CHALLENGES → NARRATIVES → SPARKS → BELIEFS → WISDOM → MODELS → FRAMES.

For each file:

1. Get the per-file detail:

   bun ~/.claude/PAI/TOOLS/InterviewScan.ts --file <NAME>

2. Check the mode:

• REVIEW mode (≥80% complete) → read the file contents to the principal first, then ask review questions
• FILL mode (<80% complete) → walk through the scanner's prompts

3. Run the conversation loop (below) 4. When the principal says "next" or "done with this one," move to the next Phase 1 file

Step 3 — Conversation loop (per file)

Review mode (for Phase 1 files at ≥80%): 1. Read the file with the Read tool. 2. Summarize what's there to the principal in 2-3 sentences. No voice here — text only. 3. Ask targeted review questions ONE AT A TIME:

• "Is <specific item> still accurate?"
• "Anything outdated to retire?"
• "Any recent thinking that belongs here but isn't captured?"
• "Anything you'd sharpen, reframe, or expand?"

4. The principal answers by voice or text. 5. If the principal wants a change, the DA writes it via Edit tool — precise old_string/new_string, preserve surrounding structure. 6. Voice-confirm only on actual changes:

   curl -s -X POST http://localhost:31337/notify \
     -H "Content-Type: application/json" \
     -d '{"message": "Updated <FILE> — captured the refinement."}' \
     > /dev/null 2>&1 &

7. Ask: "Anything else for <FILE>, or move on?"

Fill mode (for files below 80%): 1. Ask the first scanner prompt — one at a time, never a firehose. 2. The principal answers (voice or typed). 3. The DA writes the answer into the correct slot in the file — replacing TBD markers, filling empty sections, appending items. 4. Voice-confirm what got captured. 5. Next prompt. Repeat until done with this file or the principal says "next."

Step 4 — Phase transitions

After Phase 1 completes: 1. Voice: "Phase 1 done. Ready for Phase 2 IDEAL_STATE, or break here?" 2. If the principal says continue, proceed to Phase 2 top priority (usually HEALTH). 3. If the principal says stop, run final scan, voice a summary of what changed, say goodbye.

Same pattern Phase 2 → Phase 3 → Phase 4.

Step 5 — Regenerate PRINCIPAL_TELOS.md

After foundational changes, regenerate the startup summary so future sessions pick up the updates:

bun ~/.claude/PAI/TOOLS/TelosRenderer.ts 2>/dev/null || true

Rules

• One question at a time. Never dump all prompts at once.
• The principal never types schema. They speak/type the answer in their own words; the DA formats it into the file's structure.
• Always show the principal what got written before moving on. Brief voice + one line text.
• Respect stop signals. "Enough" / "stop" / "later" → save progress (state is already persistent in the files themselves), end gracefully.
• Don't ask again about filled fields. The scanner's completeness score decides what's still gap-worthy.
• Narrative dimensions stay narrative. For CREATIVE/RELATIONSHIPS, don't coerce answers into metrics — write prose that matches the principal's words.
• Target/North-Star classification. After writing a target's entries, ask once: "Is this a concrete achievable target, or a north-star orientation?" Update the type: field accordingly. (Default target.)
• Back up before multi-edit. If about to rewrite ≥50% of a file, save timestamped backup to TELOS/Backups/FILENAME-YYYYMMDD-HHMMSS.md first.

Examples

User: /interview

The DA runs InterviewScan, presents top 3 gaps, asks the principal to pick.

User: /interview --resume

The DA runs scan (same as above — state is in the files themselves, no separate session to resume).

User: /interview health

The DA skips the full scan, jumps straight to the IDEAL_STATE/HEALTH interview.

User: "next area" (mid-interview)

The DA marks current section progress via file state (already saved by Edit tool), re-scans, asks the principal to pick next highest-priority or auto-continue.

Related

• /migrate — intake content from other sources (not an interview, a one-shot classification)
• /Telos — edit a single TELOS file directly with backup
• /_PROFILE — manage PRINCIPAL_IDENTITY directly