job-discovery-research

name

job-discovery-research

description

Use when mapping career options for someone whose ideal job title isn't obvious from their background — an academic or PhD exploring industry, a consultant returning from sabbatical, or anyone whose skills transfer to roles they can't yet name. Use when ordinary job-title searches keep returning the same generic roles and you want to surface non-obvious, capability-matched positions across organizations.

Job Discovery Research

Overview

Job titles are output, not input. Most searches start from a list of titles and miss everything phrased differently — the surprising, well-fitting role you'd never think to type. This skill inverts the search: hunt on capabilities and organizations, then harvest the titles you find. The titles you didn't predict are the signal.

Two phases, in order:

1. Grill the person into a locked profile. Do this once; the profile is the spec.
2. Run a widening search loop against it. The profile stays fixed; the map of roles widens each pass — under human approval.

Do not skip Phase 1. A search without a locked profile drifts into the generic results this skill exists to prevent.

Mechanics live in reference/method.md; search vocabulary (phrase buckets, org archetypes, hop prompts, a worked example) in reference/lexicon.md.

When to use

• An academic / PhD / postdoc mapping the jump to industry or applied work
• A consultant or operator returning from a sabbatical or career break
• Anyone asking "what do my transferable skills actually transfer to?"
• When normal title-based searches keep surfacing the same obvious roles

When not to use:

• You already know the exact title you want → just search it.
• You need application help (resume, cover letter, interview prep) → use a careers/writing skill.

Phase 1 — Grill to a locked profile

Interview relentlessly (mirror grill-me: one question at a time, recommend an answer each time) until these slots are pinned and crystallized into a PROFILE.md:

• Core capabilities — the non-negotiables. If more than one is essential, the target sits at their intersection — a role hitting only one must-have is a miss, not a near-miss. Mark must-have vs. nice-to-have.
• Unfair advantage — what others can't replicate quickly regardless of talent, not merely what they do well; prioritize roles that value it.
• Hard filters — disqualifiers (domain, ethics). A miss here kills the role regardless of other fit.
• North star — what "good" means (impact, reach, autonomy, mission).
• Constraints — location, work mode, comp floor, timeline.
• Tuning dials — under-determined preferences; surface across all settings rather than pre-committing.

Ask for a resume up front (optional — without one, grill as usual). Mine it for candidate advantages — rare intersections, recurring threads, hard-to-replicate assets, buried under-claimed items — then let the grill disposition every candidate as confirmed-love / confirmed-skill / rejected before lock. Mechanics in reference/method.md.

Then lock it. Lock only once every slot is resolved (a real answer or an explicit "open/none") and you've read the profile back and gotten explicit sign-off — the full gate, including conditional checks (intersection named, vague terms sharpened, love-vs-skill tagged, mined resume candidates dispositioned), is in reference/method.md. As soon as it passes, stop asking; gilding a complete profile is over-grilling. Don't re-litigate mid-search; note candidate revisions and confirm before changing the spec.

Phase 2 — The search loop

Profile stays locked; each pass widens the map. (Seed buckets, archetypes, hop prompts: reference/lexicon.md. Harvest/rank + eval mechanics: reference/method.md.)

1. Seed capability phrases. Search posting bodies, title-agnostic. The richest finds sit at the intersection of every must-have dimension; a strong hit contains a term from every must-have bucket.
2. Org archetypes. Find the places at the intersection, then read every role they post regardless of title.
3. Adjacency hops — the novelty engine. From each strong find, follow one or two independent hops onto titles/orgs you did not seed (where people go next; who else hires this language; where alumni land) — then stop for judgment.
4. Harvest + rank. Tabulate finds; score concordance per must-have (none/some/full) and leverage (none/some/full — does the role prize the unfair advantage, or would anyone competent do?), fit = lowest across must-haves; rank by fit × distinctiveness × leverage; kill any none-on-a-must-have or hard-filter fail (a none on leverage sinks a role, but doesn't kill it). Also surface strategic meta-findings — cross-cutting patterns the table hides: constraint-unlocking org classes, on-ramps, negative findings, and consciously-off-mission bridge roles when a hard constraint forces a stopgap now.
5. Loop — on approval. Present the harvest for judgment; with the person's go-ahead and reactions, feed approved novel titles, phrases, and orgs into the next round.

Evals & quality gates

Built to go wide — breadth is the point, so don't prune speculative-but-plausible finds. Two light gates, plus the human:

• Hallucination check (the one hard gate). Confirm each org/role actually exists; flag what you can't as unverified — never silently drop it or invent a link. Do not adversarially refute fit.
• Concordance & leverage rubrics. none/some/full per must-have, fit = lowest across must-haves; leverage scored none/some/full against the unfair advantage (sinks, never kills). Coarse on purpose — three levels stay repeatable where 1–5 drifts.
• The human is the eval. 1–2 hops → stop → present for judgment; widening further needs approval. Reactions are the ground truth that steers the next round.

Tooling

Prefer full-text/boolean body search + career-page scraping — title-only boards structurally miss novel roles, and many capacity-building/curriculum/education roles live on org and sector sites, not job boards. Report cost, inputs, key/paid needs, and reliability before running. At scale, run the loop as a multi-agent fan-out with split model tiers — cheap/fast for search legwork, strong for evaluation/ranking/synthesis; see reference/method.md.

Common mistakes

• Searching a fixed list of titles → you only refind the obvious. Search capabilities; harvest titles.
• Dropping a must-have to get more hits → a pure-X role ignoring must-have-Y is a miss. Keep the intersection intact.
• Stopping at the first match → value lives one to two hops out. Always hop at least once.
• Re-litigating the locked profile mid-search → note revisions, confirm, continue.
• Fabricating orgs/links to look comprehensive → flag unverified; never invent.