job-preference-interview

job-preference-interview

To configure, extend, customize, or troubleshoot the agent itself (or understand its capabilities), use the job-search-agent skill — the operator manual.

Build the user's Job Preferences Brief at a depth they choose — from a one-line sketch to a thorough interview — the prose preferences.md that evaluate-job-fit later reads next to a posting to judge it. Follow references/voice.md for how you talk and how you present the finished brief — and don't narrate setup mechanics (no "resolving the workspace", no reference-file talk; your first words open the conversation itself).

Purpose

Produce a prose brief a model can read against a job posting and judge it qualitatively: relevant or not, and if relevant weak / moderate / strong, with plain-language reasoning. The brief is the "what I want" half of the system; evaluate-job-fit is the reader.

No numbers. Do not produce a fit score, a numeric scale, per-category point values, or category weights, and do not ask the user to rank categories against each other numerically. Importance is expressed by which bucket a factor lands in (must-have vs. strong preference vs. nice-to-have), never as math. Capture compensation, hours, and the like as plain words ("at least ~$180K base", "no regular on-call"), never as a formula.

This skill is interactive only — it asks questions and waits for answers. Never invoke it in a headless or scheduled run.

Where it writes

Find the workspace with the Discovery procedure in references/internals.md (one fact-gathering command, then its precedence rules) and never hard-code its path.

• Write the brief to <workspace>/<config.workspace.preferences_path> (default preferences.md).

• If discovery reports first_run: true and you were invoked standalone (no workspace set up yet), write to the resolved default path anyway and tell the user exactly where it went. (If onboarding owns the workspace, it tells you where to write.)

• If a preferences.md already exists, you are updating it — read it first, fill gaps, and confirm changes rather than silently overwriting.

• Put created_at: and updated_at: front-matter lines at the top, matching templates/preferences.example.md. On a new brief set both to today; on an update keep created_at and refresh updated_at to today — the home view measures staleness from updated_at:

  ---
  created_at: 2026-06-05
  updated_at: 2026-06-08
  ---
  # Job Preferences Brief
  ...
  

Choose a depth (offer all three; give the question estimate)

Before you ask anything, let the user choose how deep to go — and make clear they can start light and deepen later (a follow-up interview reads the existing brief and enriches it, never overwrites). The depth ask happens even when another skill invoked you. Onboarding's hand-off, an args string ("standard interview", "one question at a time"), or any invoker's description of this skill never counts as the user's choice — only the user's own words do (they already said "just a quick sketch" / "make it thorough" → honor that and skip the ask). The failure mode: an invoker pre-picks "standard", the depth question silently disappears, and the user never learns a one-question sketch existed.

Ask it as a closed choice (see your platform's adapter → Closed-choice question; references/voice.md → Asking questions keeps the agnostic decision). Header Depth; question: "This brief is the plain-English 'what I want' that every job posting gets judged against. How deep do you want to go? You can always come back for a deeper pass later."; options:

1. Quick sketch — "~1 question — describe what you want in a sentence or two; see matching jobs right away."
2. Standard interview — "~6–10 questions, one at a time, over what matters most."
3. Thorough interview — "~15–20 questions across every dimension, for the most precise brief."

(Standalone only — not mid-onboarding, where import was just declined — add a fourth option: Import — "already have one written down? Paste it or give me the path.")

Whatever the depth, the output is the same five-section brief (below): depth changes how much you ask, not the shape of the result. If a brief already exists, any path updates it — read it first, fill gaps, and confirm changes rather than overwriting.

Quick sketch — the fast escape hatch

For users who'd rather see jobs now than answer questions:

1. Ask once: "In a sentence or two — what are you after? (role, where, pay floor, anything that's a dealbreaker)". Take whatever they give you; don't push for more.
2. Draft the five-section brief from only what they actually said, plus safe, direct implications (e.g. an on-call red flag from "good work-life balance") — a stated role / location / pay floor becomes a Must-have, softer wants go to Strong preferences / Nice-to-haves. Don't invent preferences they didn't express; leave a section empty rather than padding it — they can deepen it later. Ask at most one follow-up, and only if a likely must-have is missing entirely.
3. Write it, show it rendered in your reply (no code fence), say in one line where it went, and tell the user plainly they can run a deeper interview anytime to sharpen it — then hand back so they can run a search.

Interview method

This is the Standard and Thorough path — same method, different coverage. Standard (~6–10 questions) works the core dimensions below and skips whatever the user says doesn't matter. Thorough (~15–20) works through every dimension with follow-ups and deliberately fleshes out all four buckets — nice-to-haves and red flags included.

Standing rules (apply to every question)

• Ask one main question at a time (a single tight, directly-related follow-up is fine). Wait for the answer before moving on. Never dump a long checklist of questions.
• Adapt to answers — let each reply decide what to probe next.
• Make vague answers concrete. When you hear "good culture", "decent pay", "work-life balance", ask a follow-up that turns it into something observable a reader could actually check against a posting ("good culture" → "small teams, low meeting load, ships weekly"; "decent pay" → "base at least ~$X").
• Make answering easy. Offer a few example options or a simple scale when it helps, and always let the user say "no preference", "skip", or "that's a dealbreaker". When a question is a genuine pick-one with 2–4 natural answers (IC vs. manager; remote / hybrid / onsite), ask it as a closed choice (see your platform's adapter → Closed-choice question); open questions stay prose (references/voice.md → Asking questions).
• Keep every message short. Don't lecture or pad.

Flow

1. Start with the user's current situation and what's prompting the search ("What's making you look now?"), then work through the dimensions below.
2. Reflect back every 4–5 questions in 1–2 sentences so the user can correct you.
3. Finish when you have enough detail or the user says they're done — then write the brief.

Dimensions to cover

Skip any the user says don't matter; add others if they come up. For each, learn what they want, how much it matters (which bucket), and what would be a dealbreaker.

1. Role — function, title, seniority, scope, day-to-day, IC vs. manager.
2. Industry / domain / mission — the kind of product, problem, or work.
3. Company — size, stage (early startup → enterprise), culture, values, reputation.
4. Compensation — base, bonus, equity, benefits, and a minimum acceptable. Capture as prose, never as math ("≥ ~$180K base; equity matters; benefits flexible").
5. Location & arrangement — remote / hybrid / onsite, geography, travel, relocation.
6. Work-life balance — hours, intensity, on-call, PTO, flexibility.
7. Growth — learning, promotion path, mentorship, skill development.
8. Team & management — team size, manager style, reporting lines, whether they manage.
9. Tools / tech stack / skills / methods used day to day.
10. Stability vs. risk — job security, funding stage, risk tolerance.
11. Hard constraints / dealbreakers — anything that's an automatic no.

Calibration (qualitative buckets, NOT weights)

At a natural point — once the dimensions are mostly covered — sort the factors into four buckets. This replaces any scoring: importance lives in the bucket, full stop.

• Must-haves / dealbreakers — absent or violated = automatic reject. Phrase each as a binary, checkable condition ("Remote within the US, or SF Bay onsite").
• Strong preferences — really want it; a strong match should hit most of these.
• Nice-to-haves — pluses, not requirements.
• Red flags — things whose presence makes a posting worse / a likely pass.

Do not ask the user to assign numbers or weights, and do not ask how categories trade off against each other numerically. If they volunteer a relative ordering, capture it in words inside the bucket.

Output: the brief

Write the prose document to the path above, with exactly these sections — references/conventions.md ("preferences.md — prose brief") owns the authoritative set (names + order); the gloss below is orientation, not a second source of truth:

• Summary — 2–3 sentences capturing the ideal role in plain language.
• Must-haves / dealbreakers — the binary filters; each phrased so a reader can check it against a posting.
• Strong preferences — the heavily-wanted, non-binary criteria.
• Nice-to-haves — the pluses.
• Red flags — anti-preferences whose presence weighs against a posting.

Every item is plain and observable — something a reader could verify against a posting's text, not an internal feeling. Skip an empty section's bullets rather than inventing filler.

End the brief with a one-line How to use this note, e.g.:

How to use this: your job-search assistant reads this brief next to each posting and judges whether it's relevant, and if so whether it's a weak, moderate, or strong match — with reasoning. No score.

After writing, show the user the brief itself — print preferences.md's body directly in your reply as normal message text (rendered markdown; no code fence, skip the front-matter lines — see references/voice.md), say in one line where it's saved, and offer to refine any section.

Import an existing brief

If the user already has a brief, accept a file path or pasted prose instead of interviewing from scratch.

1. Validate it's usable. It should be prose with at least a Summary and Must-haves.
2. If it carries a numeric rubric or category weights (a fit-score scale, per-category points, percentage weights), tell the user this system is qualitative only and offer to convert it to prose — keep the criteria, drop the numbers, and reshape into the five sections above.
3. If it's thin (missing sections, vague items), offer a few targeted enrich questions to fill the gaps — using the same one-question-at-a-time method.
4. Map its contents onto the five sections, add the created_at: + updated_at: front-matter lines, and write preferences.md at the resolved path. Show the finished brief rendered in your reply the same way (no code fence), and confirm in one line where it's saved.