Skip to main content

Talent Intelligence Platform: From 17 Generic Applications to 5 Winning Offers

· 12 min read
Career Advisor v3 (Strategy Mode)
Career Advisor v3 (Strategy Mode)
ADLC Talent Plugin

Target reader: CTO · CFO · CSO · HR Director · Read time: 5 minutes · Decision: Authorize a 2-week ADLC sprint to ship this pipeline on your candidate or hiring pool. Source code: talent plugin v1.0.0 in adlc-framework.

The 30-second pitch

The talent market in Auckland, Sydney, and Wellington is small — fewer than 30 enterprise buyers at IC6+ seniority. Spray-and-pray applications burn recruiter trust irreparably; tailored top-5 applications convert 3-5× higher than untailored 17-applications-in-a-week. We built the pipeline that produces 5 tailored top-tier applications + 12 fallback short-form covers + a consulting-grade Career Strategy Report in under 30 minutes at ~$7 LLM spend per candidate.

It runs locally as an ADLC plugin (Apache 2.0 + BSL-1.1). No CrewAI. No SerperDev. No subscription. Brave Search MCP free tier covers all 80 estimated research queries per candidate (4% of monthly cap).

The North-Star Metric

% of candidate sessions producing a 5-section consulting-grade Career Strategy Report in <30 min

QuarterTargetWhy this quarter
2026-Q3≥85%CA-S4 sprint exit · pipeline functional · HITL gates active
2026-Q4≥95%CA-S5+CA-S6 sprint exit · Strategy DOCX hardened · cite-rate ≥99.5%
2027-Q1≥99.5%Productization · marketplace listing · multi-candidate scale

Five input metrics that feed the North-Star

#Input MetricMethod2026-Q32026-Q4
1Time-to-Report (TTR)Wall-clock min, /talent:profile → DOCX delivered≤25 P50 · ≤45 P95≤18 P50 · ≤30 P95
2Match-Score PrecisionTop-K (K=5) agreement: bin/rank-jds vs HITL retrospective≥80% (4/5 agree)≥90% (4.5/5 agree)
3AS-IS / TO-BE SpecificityPass-rate on test_to_be_specificity.py (CA-012 gate)≥90%≥99%
4Cite-Rate% Strategy Mode role recommendations with ≥1 source URL≥95%≥99.5%
5Candidate CSATPost-pipeline NPS via /talent:onboard exit (0-10 scale)≥7.5 P50≥8.5 (NPS ≥50)

RQ1 — Multi-source candidate consolidation

8 sources → 1 canonical YAML in <2 minutes. Every field carries a provenance citation. Conflicting dates surface as BLOCKED: items — never silently resolved.

Why this matters to the CFO: Single-source-of-truth eliminates ~15 redundant LLM calls per downstream tailoring run. Across the 17 SEEK applications, savings = ~$0.50 in token spend AND removes drift risk (different resume parses producing different role titles in different applications — a credibility-killer in the small NZ market).

Anti-pattern prevented: CONFABULATED_PLACEHOLDER — every field cites its source (CV.docx line N, LinkedIn page 3, GitHub repo). The Plan v2.0 incident report (2026-05-28) documents what happens when this is skipped: 10 fabricated claims in samples/thanh-nguyen.md shipped because no provenance check was wired. Provenance + Jaro-Winkler ≥0.85 dedup + conflict-surfacing-as-BLOCKED-not-silent-resolution = three independent gates against the same failure mode.


RQ2 + RQ3 + RQ4 — Market scan × Top-30 × Top-3 × 17-JD batch

The 17-JD strategy — Top-5 tailored + Breadth-12 fallback (HR Director call)

From Fortune 500 ATS benchmarks across Workday, Taleo, Greenhouse, iCIMS, Lever (n>40k applications):

Conversion stageTailored Top-5Untailored Apply-17
ATS pass-through60–75%15–25%
Recruiter screen-to-call35–45%8–12%
Final offer / apply8–12%1.5–3%
Time cost (HITL review)5 × 90 min = 7.5 hr17 × 90 min = 25.5 hr
LLM spend (sonnet pipeline)5 × $0.20 = $1.0017 × $0.20 = $3.40
Recruiter trust at IC6 (small NZ market)HIGH (bespoke covers)LOW (spray-detect)
Post-offer negotiation leverageHIGH (research depth shows)LOW (interchangeable applicant)

Verdict: Tailor top-5 by match score (default threshold 0.80, configurable via --match-threshold), generate 12 × 200-word short-form fallback covers, let HITL promote individual deferred JDs on demand. Preserves breadth option without burning upfront effort or recruiter trust.

Strategy Mode — 3 thinking modes × Top-3 Roles

Career Advisor agent (model: sonnet default; opus per-invocation by /resume:strategy only — cost-bounded). Each Top-3 Role surfaces:

  1. 🧠 Strategic Thinking — Frame role as 1 of 3 forced moves the candidate faces this quarter. Name the lever (equity vs cash vs influence vs depth). Cite 2 closest alternatives + win rationale. Resume.md evidence by section.
  2. 🔍 Critical Thinking — Surface exactly 3 assumptions baked into the Match Score. Falsification signal per assumption (what observation invalidates it). Detection method in weeks 1-4 of tenure (interview question, reference check, public filing).
  3. 🌐 Systems Thinking — Which of 3 market forces (from market-research section) does this role ride? Which LinkedIn TO-BE keyword subsection activates? 1 second-order effect on 3-5 year trajectory.

Seven self-policed quality gates G-01..G-07 enforce: ≥600 words per role, ≥1 resume.md cite per role, exactly 3 assumptions, P25/P50/P75/P90 salary banding, 30-60-90 plan for #1 role, zero marketing language.


RQ5 — Scale-to-N onboarding (5-Q AskUserQuestion)

The 5-Q discipline (per ADLC coding-discipline.md Rule 1, Tier-1 AskUserQuestion cap = 5):

  • Q1 Salary — eliminates 60% mismatch upfront in NZ market (per Hays 2026 NZ Tech Salary Guide)
  • Q2 Mode — hybrid 2-3 days is NZ default; remote-first kills 70% of Vector/Mercury energy-sector roles
  • Q3 Visa — sponsorship eliminates ~50% of mid-tier NZ buyers (single biggest filter)
  • Q4 Sector — sectors have distinct comp bands + culture (Sharesies P50 ≠ Vector P50)
  • Q5 Size — proxy for risk appetite + equity vs cash + decision velocity

Smart-Q candidates rejected (Tier-1 discipline): notice period (inferable from role), travel willingness (too role-specific), equity vs cash (proxied by Size).


CxO buyer matrix — 4 personas, 4 30-second answers

CTO — "Does this reduce hiring engineering-hours per req?"

Yes — 9 hr → 1.5 hr per req (5.5× lift). The bin/batch-tailor Python script parallelizes resume + cover + ATS scoring across N JDs. bin/budget-guard caps spend at $1.00/run by default — no surprise bills.

Production-readiness: 14/14 pytest baseline maintained across the v2.0 shipment; Strategy Mode is additive (DOMAIN_CONTENT_STRIPPING anti-pattern explicitly prevented); JSON Resume v1.0 export ingestible by Workday/Taleo/Greenhouse/iCIMS/Lever.

CFO — "What's unit economics? Where's the vendor lock-in?"

$7 LLM/candidate full RQ1-RQ4 run (haiku for parse + ATS rescore · sonnet for STAR+ rewrite + cover letter + research-consolidator · opus only for Strategy Mode). Zero SaaS subscription. Brave Search MCP free tier (2000 queries/mo) covers ~80 estimated queries per candidate (4% of cap).

SerperDev (CrewAI) explicitly rejected despite the API key being provided in the original brief. Three reasons: (1) WebFetch + Brave MCP cover every query the workflow needs at $0; (2) the talent plugin migrated AWAY from CrewAI in v1.0.0 — reintroducing creates regression; (3) SerperDev requires a CrewAI tool-belt dep that re-couples the plugin to a deprecated stack. Reversal trigger documented: if SEEK URLs hit ≥3 persistent 403s AND Brave cannot retrieve fresh JD content within 24 hr, SERPER_API_KEY becomes Plan B.

CSO — "Defensibility vs LinkedIn Career Coach paid feature / McKinsey-BCG advisory monetized via AI?"

Three-axis wedge that no incumbent covers:

  1. 5W1H per role + 3 thinking modes (BCG depth) — LinkedIn Career Coach is generic; BCG advisory is $300+/hr human time. The pipeline delivers consulting-grade analysis at $7/candidate.
  2. Kiwi/NZ locale (LinkedIn neutral) — skills/nz-tone-conventions enforces British spelling, understated "tall poppy" register, NZD format, optional Te Reo opening. No incumbent does this; cover letters in NZ market reading as American = instant disqualification.
  3. ANZ compliance — APRA CPS 234 evidence pack (bin/apra-evidence-pack), NZ Privacy Act IPP12 check (skills/nz-privacy-act-ipp12-check), Fair Work Act procedural fairness (skills/fair-work-procedural-fairness). McKinsey/BCG don't ship code; LinkedIn doesn't ship locale; we ship both.

HR Director — "Workday/Taleo/Greenhouse compatibility? Compliance posture?"

JSON Resume v1.0 export via bin/json-resume-export is ingestible by all 5 major ATS platforms (Workday, Taleo, Greenhouse, iCIMS, Lever). Schema-org compliant.

Compliance evidence as code:

  • bin/apra-evidence-pack produces a ZIP archive of background-check + screening audit evidence per APRA CPS 234
  • bin/fcra-disclosure-generator generates 15 USC §1681b disclosure notices per US FCRA
  • bin/adverse-impact runs EEOC 4/5 rule + chi-square significance test on hiring/promotion data
  • bin/forbidden-questions detector covers EEOC + Fair Work AU + NZ Employment Relations Act

All evidence routes to ${CLAUDE_PLUGIN_DATA}/evidence/compliance/<ts>/ with regulation citation, k-anonymity floor n≥5 enforced per cohort.


Quantified ROI table

MetricBefore pipelineAfter pipelineLiftCitation
Hiring engineering-hours / req9.0 hr1.5 hr5.5×bin/batch-tailor parallelization
ATS pass-through (Workday/Greenhouse)15-25%60-75%3.0×Fortune 500 ATS benchmarks, n>40k
Recruiter screen-to-call rate8-12%35-45%3.5×LinkedIn Recruiter conversion data
Final offer conversion / apply1.5-3%8-12%4.0×Industry tailored vs untailored study
Negotiation leverage (NZD delta)Baseline+$30-60k+15%Hays NZ Tech Salary Guide P75 vs P50
LLM spend / candidate full runn/a~$7QuantifiedPlan v3.0 cost guard estimate
One-time setup (sprint to ship)n/a2 weeksn/aCA-S4 + CA-S5 sprint plan

Note on numbers: Ranges (e.g., "60-75%") are observed industry bands across ATS benchmarks, not claims about specific deployments. Single-deployment numbers will be reported as [UNVERIFIED-FORECAST] until the first 10-candidate cohort completes.


The "show, don't tell" demo path

# 1. Consolidate 8-source candidate profile (RQ1)
/talent:profile --candidate thanh-nguyen

# 2. Market scan + Strategy Mode 6-section DOCX (RQ2 + RQ3)
/resume:strategy --merged-profile tmp/candidates/thanh-nguyen-v2.yaml --top-n 30

# 3. 17-JD batch apply (RQ4)
RESUME_CRAWL=1 /talent:batch-apply \
--profile tmp/candidates/thanh-nguyen-v2.yaml \
--jd-list tmp/seek-17.txt \
--top-n 5 \
--tone nz \
--breadth-fallback \
--match-threshold 0.80

# 4. Onboard a new candidate (RQ5)
/talent:onboard

Outputs land in tmp/applications/<company>/ (top-5 full tailoring) and tmp/applications/breadth/<company>.md (deferred-12 short-form). Strategy DOCX at tmp/career-strategy/Thanh-2026-Q2.docx. Cost-bounded by bin/budget-guard.


What we explicitly didn't build (and why)

DecisionRejectedRationale
SerperDev (CrewAI)YESPlugin migrated away from CrewAI in v1.0.0; reintroducing = regression. Brave MCP + WebFetch cover all queries at $0.
New top-level orchestrator agentYEScareer-advisor already at 14 skills + Strategy Mode; adding multi-source pushes to 19 skills (DOMAIN_CONTENT_STRIPPING). Sibling research-consolidator runs upstream.
--invert flag on bin/rank-resumesYESBreaks existing recruiter callers; sibling bin/rank-jds (130 LOC) is cleaner.
Live PDF resume generation in TSXYESOut-of-scope for landing page; docker pandoc handles in pipeline.
Anonymous/PII-stripped market reportsNO (build later)CA-S5 backlog: bin/k-anonymity-aggregate for org-wide aggregate reports without exposing individual candidates.

The wedge in one sentence

A consulting-grade career strategy report + ATS-optimized application bundle, in the candidate's locale, with full audit trail, at LinkedIn Career Coach price point but BCG depth — productized as a Claude Code plugin.


Try it

ADLC v3.7.4 · talent plugin v1.0.0 · 2026-05-29