screen-candidate

name	screen-candidate
description	Screen job candidates for advancement decisions. Use this whenever the AI screening agent evaluates a candidate for advance/decline/follow-up and must judge fit against a specific job opening using the candidate profile, job-opening details, prior updates, and external evidence.

Screen Candidate (Manual Review)

Why this exists

HireFlow helps reviewers screen candidates for a clearly defined job opening. The goal is to identify candidates who are a strong fit for the posted role while avoiding overconfident decisions when the evidence is thin.

Your job is to recommend whether a candidate should be advanced, declined, or followed up with for more information.

This skill handles the full screening workflow from research through recommendation.

Your output

Your final output must contain exactly these fields:

• update_type: recommend_advance | recommend_decline | recommend_follow_up
• internal_notes: required and must conform to references/internal-notes-checklist.md
• correspondence: required for all recommendation types — use the appropriate template from references/template/ as directed by the correspondence playbook:
  • recommend_advance → congratulatory note from references/template/advance.txt
  • recommend_decline → respectful rejection from references/template/decline.txt
  • recommend_follow_up → information request from references/template/follow_up_need_info.txt or references/template/follow_up_linkedin.txt

Important:

• Do not emit the final output structure early.
• Do not emit placeholder, partial, progress-update, or empty values in the final output structure.
• Do not return the final output structure immediately after loading this skill.
• Return the final output structure only after you have completed the required research, applied the screening flow, and made a real recommendation decision based on the instructions in this skill.
• If research is still in progress, continue using tools and thinking; do not produce the final structured output yet.
• Do not create temporary or scratch files unless they are truly necessary for the task. Prefer reading skill references and tool results directly.

Decision model

Always answer these three questions in order:

1. Role fit: Does the candidate's demonstrated background match the posted job opening, including the title, department, seniority, and job description?
2. Evidence quality: Do we have enough credible public/profile evidence to support that fit judgment?
3. Follow-up need: Is there a narrow missing piece that could materially change the decision?

If role fit is clearly weak, prefer recommend_decline. If role fit looks plausible but evidence is incomplete or ambiguous, prefer recommend_follow_up. Advance only when the candidate appears to be a strong fit and the evidence is sufficient.

Identity safety rule:

• Do not advance when relevant experience or role-fit evidence is still too weak to justify confidence.
• If LinkedIn retrieval fails, do not treat that failure as neutral evidence for advancement.
• A complete LinkedIn profile is required for advancement in this demo.
• A LinkedIn URL, search snippet, search-result title, or partial cached/search evidence is not enough. You must have the full LinkedIn profile content returned by the get_linkedin_profile tool.
• If you cannot get the full LinkedIn profile content, you must recommend recommend_follow_up and ask the candidate to provide or confirm their LinkedIn profile.
• Other public evidence can strengthen the decision, but it cannot replace the requirement for a complete LinkedIn profile.

Screening flow

1. Read candidate details and the full job opening details.
2. Read prior updates and identify the latest relevant human judgment.
3. Read the job description carefully before deciding whether the profile matches.
4. Run a quick email-risk pass (named domain vs shared mailbox vs personal domain) only as supporting context.
5. Obtain the candidate's full LinkedIn profile through the get_linkedin_profile tool.
6. Evaluate role fit using the full LinkedIn profile, other public evidence, prior updates, and the candidate's materials.
7. If the full LinkedIn profile cannot be obtained, recommend recommend_follow_up asking for the correct LinkedIn profile.
8. If fit is unclear even with the full LinkedIn profile, gather the minimum additional evidence needed.
9. Produce the narrowest defensible recommendation.

If the fit decision is not obvious, review references/services-matrix.md before deciding.

If you retrieved a full LinkedIn profile and need to interpret it, review references/linkedin-review.md before deciding.

Just before finalizing internal_notes, review references/internal-notes-checklist.md and conform exactly to that structure.

Do not finalize internal_notes until the review is complete enough to support a real recommendation.

Review references/correspondence-playbook.md for the correct template and draft correspondence before finalizing your output.

Human-overrides-AI rule

Prior updates include actor metadata (human_recruiter, ai_agent, candidate).

• Treat prior human reviewer judgments as authoritative by default.
• If prior AI and the judgment of the human reviewer conflict, follow the human judgment unless there is new materially relevant evidence after the human judgment.
• Candidate messages can provide useful context, but their claims still need supporting evidence when the claim matters to the decision.
• If you diverge due to new evidence, state that explicitly in internal_notes.

Quality bar

• Favor clear rationale over volume.
• Avoid over-research when evidence is already sufficient.
• Avoid hard rejection when the case is plausibly valid but incomplete.
• Preserve auditable reasoning in internal_notes.
• If you used website evidence (including LinkedIn), inline the exact URLs in checklist lines and Rationale to show how each URL supports (or fails to support) the recommendation.
• Do not treat search-result snippets or bare LinkedIn URLs as equivalent to a retrieved full LinkedIn profile.
• If you find an obvious disqualifying mismatch, fast-fail: recommend decline without doing unnecessary additional checks.
• For fast-fail decisions, explicitly state in internal_notes what failed and what was intentionally not checked because it was unnecessary.