Candidate Screening

1. Load active JD + criteria from role pack.
2. Pull target candidates.
   • For Indeed pipelines: indeed-cli sync:candidates ... --download
   • Default batch size: 20 candidates per pass (unless user specifies otherwise).
3. Extract resume text.
   • Use bundled script scripts/extract_resume_text.sh.
4. Enrich missing signals.
   • LinkedIn verification must use browser tool.
   • Do not treat blocked/unavailable sections as confirmed evidence.
5. Score and rank.
   • Apply hard gates first.
   • Score competencies 1-5 with concise rationale.
   • Sort highest to lowest.
6. Report with provenance and inference labeling.
7. Persist notes for continuity (required).

Notes + continuity (required)

Store notes inside the selected role-pack under:

• notes/<YYYY-MM-DD>/batch-<NN>/

Current example for this role:

• notes/2026-02-28/batch-01/

For each batch, create:

1. batch-summary.md
   • ranked list, final decisions, key callouts, unresolved questions
2. candidate-notes.md
   • one section per candidate with detailed analysis
3. evidence-log.md
   • exact source trail per candidate:
     • resume path(s)
     • LinkedIn URL(s) checked via browser
     • any additional web sources used
     • what was inferred vs explicit

Keep these notes concise but decision-grade so future runs can continue without re-reading all raw files.

Output format (default)

• Role + rubric file paths used
• Ranked list (highest score first)
• Per candidate:
  • Name, score, decision
  • Hard gates (PASS/FAIL/UNCERTAIN)
  • Top strengths
  • Top risks
  • Evidence Source:
    • Resume only
    • Resume + LinkedIn (browser)
    • Resume + other web research
  • Inference Notes:
    • What was inferred
    • What was explicit
• Final recommendation summary

See references/output-template.md for exact layout.

Guardrails

• Never fabricate missing facts.
• If EN/FR or Montreal in-person is not clearly verified, mark UNCERTAIN.
• If user gives threshold overrides, apply and restate them.
• LinkedIn evidence should come from browser workflow, not web snippets alone.