Resume Optimization（简历优化 · 环节二）

Overview

Stage two of the three-stage resume pipeline (diagnosis → optimization → report).
The deliverables are the final resume in three formats: resume.html + resume.pdf
+ resume.docx.

Entry Modes & Artifact Handoff

Default output directory is ./resume-output/; a path the user explicitly names takes
priority. Artifact lookup order: user-named path → ./resume-output/ → ask the user.

• Entry A — continuing after diagnosis. Read ./resume-output/diagnosis.json and
  build the rewrite work list from its issues / risks, its jd_match items marked
  missing/partial, its first_scan.missing facts, its timeline.concerns, and the
  user's answers to its gaps questions; calibrate evidence depth and wording against
  its seniority; read ./resume-output/before.* for source content. Recover the
  gaps answers from ./resume-output/gap-answers.md (cross-session persistence
  contract — a numbered markdown list with each question and its answer or （待补充）);
  if the file is absent or incomplete, fall back to the gaps list in diagnosis.json
  and re-ask any question whose answer is still missing.
• Entry B — direct entry, existing resume. If the user explicitly skips
  diagnosis, proceed, but first save the original resume as
  ./resume-output/before.<ext> so stage three can compare. If the user has not
  said skip and no diagnosis.* is found, suggest once: "建议先用 resume-diagnosis
  做环节一诊断，要先诊断吗？" — if declined, proceed and do not ask again.
• Entry C — from scratch, no existing resume. No before.*; stage three will use
  report "mode": "after_only".

Write resume.html / resume.pdf / resume.docx. Never overwrite an existing
before.*.

Quick Routing

• A generic "优化简历/改简历" request on an existing resume, without an explicit
  diagnosis skip → start with resume-diagnosis.
• User wants only a before/after comparison report → use resume-report.
• If the target role is missing and cannot be inferred, ask one concise question.

Runtime Compatibility

• Use only relative paths from this skill folder when loading references, templates, or scripts.
• Treat agents/openai.yaml as optional Codex/OpenAI UI metadata; Claude Code can ignore it.
• Avoid relying on product-specific tools unless the user explicitly asks for that runtime's integration.
• Prefer shell commands that work in both Claude Code and Codex: python3, local scripts, and local file paths.

Required Workflow

1. Determine the entry mode and locate (or create) the artifacts above.
2. Build an input inventory; apply the Minimum Input Standard in references/hr-scorecard.md.
3. Read references/hr-scorecard.md and references/rewrite-guide.md.
4. When a target role is named, a JD is provided, or the role can be inferred, read references/role-playbooks.md.
5. Read references/geo-html-resume.md before generating HTML or optimizing for ATS/GEO.
6. If diagnosis.json exists, convert each issue/risk into a concrete edit and land it. Extend the work list with each jd_match item marked missing/partial and with the user's answers to the diagnosis gaps questions: read ./resume-output/gap-answers.md to recover answers across sessions (fall back to diagnosis.json's gaps and re-ask any question still unanswered). Cover JD gaps only with real evidence from the user's background (anything still uncovered goes into the delivery summary's gap list), and write each answered gap into the relevant section. Surface each first_scan.missing fact in the top third (headline / summary / contact block) when the source material or a gap answer supplies it, and resolve each timeline.concerns entry not already covered by a risk with factual context — never invented history. Calibrate evidence depth and wording against the diagnosed seniority (Seniority Calibration in references/hr-scorecard.md).
7. Rewrite each section with evidence (action + scope + method + impact) and optimize keywords naturally (no stuffing).
8. Generate ./resume-output/resume.html from assets/resume-template.html — replace sample data, meta tags, JSON-LD, section headings, and all visible content.
9. Run python3 scripts/check_resume_html.py resume.html and fix every failure before claiming the artifact is ready.
10. Export by default: python3 scripts/export_resume_pdf.py resume.html resume.pdf and python3 scripts/export_resume_docx.py resume.html resume.docx, then verify both exist. If no Chrome/Chromium is available, deliver HTML + DOCX, say so, and include the PDF export command. Re-export loop (hard rule): any subsequent edit to resume.html — whether driven by user feedback, a missed check_resume_html.py failure, or a late polish pass — MUST be followed by re-running both export commands before the next delivery. Never hand the user HTML plus a stale PDF/DOCX. Converge all HTML edits first, then re-export once, rather than re-exporting after every single-line change; but the moment you deliver or claim "done", all three files must reflect the same content.
11. Stage gate. After delivering the three resume formats, ask the user: 「优化完成。是否进入环节三「复盘」，生成优化前后对比报告？」 Do not generate the report until the user confirms. On confirmation, continue with the resume-report skill and tell it ./resume-output/ holds before.*, diagnosis.json, and resume.*.
    • Pre-authorization: if the user asked up front to complete all three stages at once, treat that as confirmation; at the boundary still report the artifacts and state you are proceeding to stage three.

HR Writing Rules

• Lead with fit for the target role, not biography.
• Prefer strong nouns and verbs over adjectives: "reduced review time by 30%" beats "excellent communication skills".
• Use real numbers when provided; never invent metrics. If a metric is missing, ask, estimate only when explicitly authorized, or use scale/proxy facts.
• Convert duties into outcomes. Preserve a short responsibility phrase only when the role context would otherwise be unclear.
• Remove weak claims: "familiar with", "responsible for", "participated in", "good at", "hardworking", and unsupported self-evaluations.
• Keep one resume optimized for one target. If the user has multiple targets, create variants.

HTML Artifact Rules

• Keep the resume as semantic HTML, not an image or canvas.
• Preserve mobile readability: no fixed-width content wider than the viewport, no tiny text, no horizontal scroll.
• Preserve PDF export quality: include @page, @media print, page-break controls, and no screen-only critical content.
• Include GEO signals: meaningful <title>, meta description, meta keywords when useful, schema.org Person JSON-LD, machine-readable contact details, and target-role keywords in natural visible text.
• Avoid hidden keyword blocks, white-on-white text, fake roles, fake skills, or unverifiable claims.
• Three-format sync: resume.pdf and resume.docx are derived from resume.html, not siblings. After step 10, any HTML edit must be followed by re-running both export commands before the next delivery — never ship a newer HTML with an older PDF/DOCX.

Output Shape

• The resume in three formats: resume.html + resume.pdf + resume.docx (HTML + DOCX if PDF export is impossible).
• A short summary: role-specific keyword/evidence gaps, assumptions and missing facts, and the check_resume_html.py result.
• Every artifact path.
• The stage-gate question (unless pre-authorized).