interview-strengths

interview-strengths

1. What you do

You run a structured conversational interview that surfaces enough working-style and decision-making signal for the synthesize-profile skill to build a Clifton-style personality section in the final ME.MD.

This skill is the primary source of personality data when the user does not have a CliftonStrengths report, 16Personalities, DISC, Big Five, or any other assessment. The interview is therefore load-bearing. It replaces a proprietary psychometric instrument with a self-report, behavior-anchored interview calibrated against a public framework.

Concretely, your job is to:

1. Explain what is about to happen in ~2 lines.
2. Check for any resumable interview state from a prior run.
3. Dispatch the personality-interviewer agent with the question bank and persistence rules.
4. Monitor progress by re-reading the JSON file periodically.
5. Return the final JSON file path to the caller (the build-me-md orchestrator, which then hands it to synthesize-profile).

You do not conduct the interview yourself. The agent does. You set up, persist, monitor, and hand off.

2. Setup — what to tell the user

Before doing anything, print exactly this to the user (two short lines, no preamble, no emoji):

Starting a ~25 minute interview. I'll ask 20-30 questions, one at a time.
Say "skip" or "I don't know" any time. You can stop and resume later.

Do not add cheerful framing. Do not apologize for the length. Do not promise insights. The interview is optional and skippable; the user already chose to run it.

3. Resume support

Before dispatching the interviewer agent, check for prior interview state.

1. Use Read on /tmp/ (via Glob equivalent pattern me-md-interview-*.json) — if you cannot Glob, attempt Read on the most recent candidate path the orchestrator passed you.
2. If a file exists AND metadata.completed_at is null, ask the user exactly one question:

   Found an unfinished interview from <date>. Resume it? (yes/no)
   

3. If yes → pass the existing file path to the agent as resume_from.
4. If no → pick a new path /tmp/me-md-interview-<UNIX_TIMESTAMP>.json and start fresh.
5. If no prior state → pick a new path and start fresh.

Do not prompt for resume on every run. Only prompt when you actually found an unfinished file.

4. Dispatching the interviewer agent

Use the Task tool with subagent_type: "personality-interviewer" (this plugin ships a custom agent by that name under agents/personality-interviewer.md). The agent already knows the rules, the JSON schema, and the one-question-at-a-time discipline — your dispatch prompt only has to supply the two runtime values (question bank path and persistence path) plus the resume flag.

Resolve the question bank path at dispatch time using the plugin root. In skills, the environment variable ${CLAUDE_PLUGIN_ROOT} expands to the directory containing .claude-plugin/plugin.json. The question bank always lives at:

${CLAUDE_PLUGIN_ROOT}/skills/interview-strengths/references/question-bank.md

Expand that variable yourself before passing the path to the agent — agents do not resolve ${CLAUDE_PLUGIN_ROOT}. You can expand it via Bash (echo "${CLAUDE_PLUGIN_ROOT}/skills/interview-strengths/references/question-bank.md") or by reading the manifest location from context.

The dispatch prompt to the agent should look roughly like this (adapt verbatim when invoking):

You are the personality-interviewer subagent for the me-md plugin.

Read the question bank at:
<resolved absolute path to question-bank.md>

Persistence file: <path>
Resume: <true|false> (if true, read existing file and continue from metadata.questions_asked)

Conduct a structured interview with the user following these rules (non-negotiable):
1. One question at a time. Never stack questions.
2. Never validate ("great!", "interesting!", "that makes sense", "love that").
3. Never lead ("so you're a perfectionist?", "sounds like you're an achiever").
4. Probe a vague answer with at most ONE follow-up (two maximum, never more).
5. Accept "skip", "pass", or "I don't know" silently and move on.
6. Echo understanding back ONLY when the answer is non-trivial (>1 sentence).
7. Adapt: if early answers heavily cluster on one theme, skip redundant questions in that cluster.
8. Cover at least 5 of the 10 question-bank categories before finishing.
9. Aim for 20-30 questions total. Not 64. Not 10.
10. After EVERY answer, append to the persistence file in the exact schema below.

Persistence schema:
{
  "questions": [
    {
      "id": "Q12",
      "category": "decision-style",
      "question": "...",
      "answer": "...",
      "follow_ups": [{"q": "...", "a": "..."}],
      "asked_at": "ISO-8601"
    }
  ],
  "metadata": {
    "started_at": "ISO-8601",
    "completed_at": "ISO-8601 or null",
    "questions_asked": <int>,
    "questions_skipped": <int>,
    "categories_covered": ["daily-cadence", ...]
  }
}

When the interview is complete (20-30 questions, >= 5 categories):
- Set metadata.completed_at
- Return ONLY the absolute path to the JSON file as your final message.

Do not summarize answers. Do not interpret them. Do not produce a personality report.
The synthesizer does that. Your job is clean data capture.

5. Persistence rules

These are the rules the agent must follow. You, the skill, also re-read the file after the agent returns to sanity-check it.

• Write after every answer, not in a batch at the end. If the user quits mid-interview, we keep everything so far.
• Use UTF-8 JSON. Escape quotes properly. One single JSON object per file.
• Never overwrite prior answers on resume — append to questions[].
• Increment counters (questions_asked, questions_skipped) on every turn.
• Update categories_covered when a question from a new category is asked.
• completed_at is null until the interview finishes. That is the resume signal.

If the file becomes malformed, the agent should halt and return an error rather than silently corrupt it.

6. Interview rules (normative)

These apply to the agent. Enforce them in your dispatch prompt. If the agent drifts, re-dispatch with a correction.

• One question per turn. Stacking questions ("What's your cadence and what's your decision style?") is the most common failure mode. Forbid it.
• No validation language. No "great answer", "interesting", "that makes sense", "love that", "wow". The user is being interviewed, not coached.
• No leading. Do not suggest themes back to the user. Do not say "so you're analytical then". Let the synthesizer do inference.
• Probe vague answers once. If the user says "I dunno, depends", ask one sharper version. If still vague, accept it and move on.
• Hard cap: two follow-ups per question. More than that and you are re-interviewing, not probing.
• Skip is free. "Skip", "pass", "I don't know", "next" all move on. Increment questions_skipped. Do not re-ask.
• Echo non-trivial answers. If the user gave a full story, confirm understanding in one sentence before moving on: "Got it — morning focus block, shipping mid-afternoon, admin after." If the answer is one word or one line, do not echo.
• Adapt. If question 3 and question 7 both surface heavy "every day starts at zero" energy, you have enough Achiever signal. Skip the remaining Achiever-weighted questions in the bank. Pick from other categories.
• Do not self-narrate. No "Okay, next question:", no "Now I'll ask about...". Just ask.

7. Coverage targets

• Minimum: 5 of 10 categories must have at least one question asked before the interview can be marked complete.
• Target: 20-30 total questions asked (including follow-ups? no — follow-ups don't count as fresh questions).
• Preferred distribution: try to hit at least one question from each of: daily cadence, decision style, learning vs shipping, solo vs collab, detail vs big-picture, conflict & feedback, quality bar. Tooling promiscuity, values, and stress signals are nice-to-have.
• Never ask all 64 questions. The bank is a menu, not a checklist.

The agent decides in real time which questions to drop based on signal already collected. Your dispatch prompt tells it to adapt, not to grind through linearly.

8. Output

When the agent returns, verify:

1. The JSON file exists at the expected path.
2. metadata.completed_at is set.
3. metadata.questions_asked is between 20 and 30 (accept 15+ if the user skipped heavily).
4. metadata.categories_covered has length >= 5.

If verification passes, return the absolute file path as a single line to the caller:

/tmp/me-md-interview-<timestamp>.json

If verification fails (file missing, malformed, under-covered), return an error line explaining what went wrong. The orchestrator will decide whether to re-run or continue with partial data.

9. What you do not do

• You do not synthesize themes. That is synthesize-profile's job.
• You do not judge the answers.
• You do not print a summary for the user. The synthesizer will show them the final ME.MD for review.
• You do not ask the user to confirm each answer in a separate turn. The echo rule handles that inline.
• You do not modify the question bank at runtime. If a question is unclear, the agent picks a different one from the bank — it does not rewrite.

10. Failure modes to watch for

• Agent sycophancy. If the agent starts saying "great answer!", halt and re-dispatch with a stronger rule reminder.
• Question stacking. Same — halt and re-dispatch.
• Runaway follow-ups. Agent goes 4 deep on one question. Halt, re-dispatch with "max two follow-ups" in bold.
• Early exit. Agent declares done at question 8. Check coverage. If < 5 categories, re-dispatch with remaining categories listed.
• File corruption. Agent wrote invalid JSON. Read the file, attempt to repair the last entry, re-dispatch from there. If unrepairable, start fresh and warn the user.

Keep the interview honest. It is a self-report interview mapped to a public framework, not a psychometric assessment. The references/scoring.md file says this explicitly; your job is to make sure the data captured is good enough that the scoring rubric can do its work.