Evaluate

Scoring rules: shared/references/evaluation.md · Priority hierarchy: shared/references/priority-hierarchy.md

Turns one posting into a decision: apply, skip, or a question for the user. Replaces High/Medium/Low with a 1–5 score, because "Medium" does not tell anyone whether to spend an evening.

Step 0: Prerequisites

Per shared/references/prerequisites.md: CV and profile both required. Scoring a role against an unknown candidate produces a number that means nothing — if there is no profile, say so and run /wingman:setup interview instead.

Step 1: Get the posting

• URL → node "${CLAUDE_PLUGIN_ROOT}/tools/posting.mjs" <url> first. It reads the JSON-LD that Personio and softgarden both publish, and it exits 3 rather than hand you a page that returned HTTP 200 with no job in it. Only if it exits 3 for "no structured data" do you fall back to a scoped selector per shared/references/web-extraction.md. Never get_page_text.
• Pasted text → use it directly.
• last → the most recently modified job folder.

If the page cannot be read, stop and ask the user to paste it. Never evaluate from a job title alone, and never fill gaps by guessing what an employer probably wants.

If the extractor reports the posting expired or closed, say so plainly and stop — that is a finished answer, not a setback. Offer to look for the role on the employer's current vacancies page. Any WARN it prints (a validThrough in the past, a posting open for many months) goes into Block G, where it is reported and left score-neutral.

Save to DATA_DIR/jobs/[company-slug]-[date]/posting.md, employer URL at the top.

Step 2: Language decision

Run shared/references/language-decision.md and write language.md. Do this before any document work, because it changes every downstream artifact. If language.md already exists, use it silently.

Step 3: Build the report

Write blocks A–H per shared/references/evaluation.md:

• A Role — title exactly as advertised, employer type, contract, Tarif band, Kennziffer, location, remote policy
• B Requirement match — one row per requirement, each marked quoted / inferred / gap. Never write a row you cannot source.
• C Language — the decision and any level gap
• D Work authorization — per work-authorization.md, and do not flag a permit holder as blocked by a "no sponsorship" line
• E Compensation — Jahresbrutto including the 13th month, Urlaub, bAV/VWL, versus researched market rate. Name sources and the date.
• F Employer signal — kununu rating and what reviews actually say, Betriebsrat, Zeitarbeit detection, Impressum, culture tells from the ad
• G Legitimacy — ghost job, repost, scam. Score-neutral, reported separately.
• H Strategy — only at 4.0+

Step 4: Score

One holistic 1–5 across requirement fit, level fit, compensation, feasibility, and employer quality. Apply the hard-blocker caps.

Do not inflate to be encouraging. A tool that scores everything 4+ is one the user stops trusting, and it costs them evenings. Do not deflate to look rigorous either.

Step 5: Present and gate

Save the full report to evaluation.md. Show a compact summary: score, two strongest matches, two real gaps, any blocker, recommended action.

Then gate on the score per shared/references/interaction.md:

Append a row to tracker.md per shared/references/tracker.md. Set Board — where this posting came from (Arbeitsagentur, StepStone, Personio, LinkedIn, Interamt, Direkt, Netzwerk); if the user pasted a URL with no context, infer it from the host and say which you recorded. patterns cannot compute response rate by board from a blank column, and the row is only cheap to fill now.

Response Format

1. Score — the number, with the one-line reason
2. Strengths / Gaps — two each, sourced
3. Blockers — language, authorization, dealbreakers, or none
4. Legitimacy — only if something is off
5. Recommendation — apply / ask / skip, with the next command

Permissions

{ "permissions": { "allow": [
  "Read(~/.wingman/**)", "Write(~/.wingman/**)", "Edit(~/.wingman/**)",
  "mcp__claude-in-chrome__*"
] } }