joonyeonglim/realistic-job-market-research

Realistic Job Market Research

Purpose

Reproduce a complete two-stage workflow: declared-source census and artifact audit first, personalized realistic review second. Distinguish low-cost application preparation from the stricter decision to join an employer.

Before either mode, read and validate the private profile described in references/personalization.md. Default path: ~/.config/realistic-job-market-research/profile.json.

Resolve supporting scripts from this skill directory, never from the user's project cwd. Claude Code can use ${CLAUDE_SKILL_DIR}; in Codex, use the absolute directory containing this SKILL.md.

Workflow

1. Freeze the candidate contract from the local profile and verify referenced resume hashes. Do not paraphrase a mutable resume from memory.
2. Choose a mode. census owns declared public collection and the filterable ledger. review starts from an audited ledger or named shortlist. update compares prior IDs and bodies with current pages.
3. For census, read references/census-runbook.md, initialize a new immutable run, attempt all 29 registered sources, preserve blocked or failed zero-row artifacts, sync terminal states, build the manifest and dashboard, then run both executable gates.
4. For review, recheck every selected role on its current detail page. The exact body and source ID outrank title, tags, and search snippets.
5. Extract mandatory and preferred requirements, duties, employment, exact location, work policy, compensation, deadline, and hiring steps. Use UNKNOWN when absent.
6. Resolve employer identity before finance. Match every mandatory requirement to profile evidence as confirmed, transferable, missing, or unknown.
7. Read references/scoring-model.md, then calculate JD Match Score, Opportunity Score, and Evidence Confidence with scripts/score_review.py. Calculate raw weighted values at full precision, apply non-compensatory ceilings before bands and ranking, and report sensitivity profiles.
8. Make two decisions: PREPARE | CONDITIONAL | DROP for application effort and PASS | HOLD | NO_GO for the employer or offer. Scores inform these states but never replace them.
9. Adversarially recheck leaders using references/pitfalls.md, including degree floors, customer-facing work, language, mandatory frameworks, scale, hidden tests, stale finance, and founder-commitment signals.
10. Produce coverage, scored shortlist, component scores, sensitivity changes, hard exclusions, corrections, claim-level sources, unknowns, and resume actions. Validate machine-readable output.

Modes

• Census: run <skill-dir>/scripts/run-census.mjs --run-dir /absolute/run/path with Node.js.
• Named shortlist: deeply verify roles supplied by the user.
• Audited ledger: select a small review set; unreviewed rows are not exclusions.
• Update correction: label genuinely new, reposted, previously missed, closed, and materially changed roles separately.

Rules

• Prefer official career pages and exact job details. A search result is discovery evidence, not proof that a role is active.
• Never call a query-limited, blocked, or failed source complete. Never synthesize rows from provider totals.
• Preserve conflicting facts with both sources; do not average them.
• Do not infer salary, work policy, degree requirements, finance, or hiring steps.
• Do not express unsupported hiring probabilities as percentages. Use reasoned high, medium, or low fit labels.
• Never present a score without its components, numeric calculation trail, evidence confidence, model version, weights, and active caps.
• Do not store recruiter names, emails, phone numbers, private messages, credentials, or confidential employer data.
• Do not apply, save, message, edit a profile, or contact anyone without explicit authorization for that action.
• Keep the final shortlist small. More rows are not a better result.

References

• references/census-runbook.md — read for collection, resume, and artifact transitions.
• references/source-registry.md — 29-source scope and completeness contract.
• references/raw-schema.md and references/verification-gates.md — read before build or audit claims.
• references/evidence-contract.md — read before finance, identity, status, or decision classification.
• references/review-schema.md — read when producing or validating machine-readable review JSON.
• references/scoring-model.md — research basis, formulas, weights, caps, and sensitivity analysis.
• references/personalization.md — local profile and public/private boundary.
• references/authenticated-sources.md — read before any personalized browser source.
• references/pitfalls.md — adversarial checks learned from failed scans.
• references/reproducibility.md — frozen reference-run parity evidence.