--- name: cv-import description: "Extract a PDF/DOCX CV, enhance it with LLM, and transform it into the qwik-resume-editor JSON schema (Resume v2.0.0). Uses markitdown if available, falls back to LLM extraction." allowed-tools: [ Read, Write, Bash ] --- # CV Import — PDF to Resume JSON Import any CV/resume PDF into the qwik-resume-editor JSON format in three stages: **Extract → Enhance → Transform**. ## Usage ``` /cv-import /path/to/cv.pdf # writes src/data/default-resume.json (default) /cv-import /path/to/cv.pdf --output resume.json # write to a custom file instead ``` ## Stage 0 — Validate Input 1. Confirm the file path was provided. If not, ask the user: `"Please provide the path to your CV PDF."` 2. Check the file exists: ```bash test -f "" && echo "exists" || echo "not found" ``` 3. Accept `.pdf`, `.docx`, `.doc`, `.txt`. Warn (but proceed) for other extensions. ## Stage 1 — Extract Text ### 1a. Check for markitdown ```bash markitdown --version 2>/dev/null && echo "MARKITDOWN_OK" || echo "MARKITDOWN_ABSENT" ``` ### 1b-A. If markitdown is available ```bash markitdown "" ``` Capture the full stdout as the extracted markdown. markitdown handles multi-column PDFs and strips headers/footers well. ### 1b-B. If markitdown is absent — LLM extraction Use the `Read` tool to open the PDF directly. Claude can read PDF files natively. If that fails (non-text PDF), instruct the user: ``` markitdown is not installed. Install it for best results: pip install markitdown # or: pipx install markitdown Falling back to native PDF reading via Claude… ``` After reading, extract **all visible text**, preserving approximate section order. Do not summarise at this stage — get everything. ## Stage 2 — Enhance After extraction, use LLM reasoning to: 1. **Identify sections** — find headings that correspond to: Summary/Objective, Experience/Work History, Education, Skills, Languages, Certifications, Awards/Honors, References. 2. **Normalise dates** — convert any format (`Jan 2020`, `01/2020`, `2020`) to `"YYYY-MM"`. Use `""` (empty string) for "Present" / "Current" / "Now". 3. **Clean description text** — convert bullet points to HTML ``. Wrap paragraphs in `

…

`. Preserve bold/italic if detectable. 4. **Infer contact types** — classify each contact detail: - email address → `type: "email"` - phone number → `type: "tel"` - URL (http/https, LinkedIn, GitHub, portfolio) → `type: "url"` - plain text (city, country) → `type: "string"` 5. **Infer language levels** — map free-text proficiency to the enum: `Basic | Elementary | Intermediate | Conversational | Proficient | Advanced | Native` 6. **Score expertise** — if expertise/competency levels are present, map to 0–100. 7. **Flag uncertain fields** — add a `_review` key next to uncertain values with a note, e.g. `"_review_start": "date was '2020' — assumed January"`. ## Stage 3 — Transform to Resume JSON Produce a valid `Resume` object (schema version `"2.0.0"`). ### ID generation Use this pattern for every `id` field: `_<7 alphanumeric chars>`. Generate sequential IDs: `c_0000001`, `sec_0000001`, `exp_0000001`, `edu_0000001`, `grp_0000001`, `lng_0000001`, `cer_0000001`, `awd_0000001`, `ref_0000001`. ### JSON schema ```jsonc { "version": "2.0.0", "header": { "name": "Full Name", "title": "Professional Title / Headline", "contacts": [ // Order: email, phone, LinkedIn, website, location last { "id": "c_0000001", "type": "email", "label": "you@example.com" }, { "id": "c_0000002", "type": "tel", "label": "+1 234 567 8900" }, { "id": "c_0000003", "type": "url", "label": "linkedin.com/in/you", "href": "https://linkedin.com/in/you" }, { "id": "c_0000004", "type": "string", "label": "City, Country" } ] }, "theme": { "paletteId": "orange-navy" }, "sections": [ { "id": "sec_0000001", "type": "summary", "title": "Summary", "visible": true, "data": { "text": "

Professional summary as HTML…

" } }, { "id": "sec_0000002", "type": "experience", "title": "Experience", "visible": true, "data": { "items": [ { "id": "exp_0000001", "title": "Job Title", "company": "Company Name", "location": "City, Country", "start": "2021-03", "end": "", "description": "
  • Achievement or responsibility
" } ] } }, { "id": "sec_0000003", "type": "education", "title": "Education", "visible": true, "data": { "items": [ { "id": "edu_0000001", "degree": "Bachelor of Science in Computer Science", "school": "University Name", "start": "2015-09", "end": "2019-06" } ] } }, { "id": "sec_0000004", "type": "skills", "title": "Skills & Knowledge", "visible": true, "data": { "groups": [ { "id": "grp_0000001", "label": "Programming Languages", "skills": ["Python", "TypeScript", "Go"] } ] } }, { "id": "sec_0000005", "type": "languages", "title": "Languages", "visible": true, "data": { "items": [ { "id": "lng_0000001", "name": "English", "level": "Native" }, { "id": "lng_0000002", "name": "Mandarin", "level": "Conversational" } ] } }, { "id": "sec_0000006", "type": "certifications", "title": "Licenses & Certifications", "visible": true, "data": { "items": [ { "id": "cer_0000001", "name": "AWS Solutions Architect", "issuer": "Amazon Web Services" } ] } } // Only include section types present in the CV. // Available types: summary | experience | education | skills | languages | // certifications | awards | expertise | references ] } ``` ### Section ordering heuristic Default order (omit sections not found in the CV): 1. `summary` 2. `experience` 3. `education` 4. `skills` 5. `expertise` 6. `languages` 7. `certifications` 8. `awards` 9. `references` ### Skills grouping heuristic If the CV lists skills as a flat list, group by inferred category: - Programming Languages - Frameworks & Libraries - Tools & Platforms - Databases - Soft Skills / Other If no clear grouping exists, use a single group `"label": "Skills"`. ### Expertise section Only emit an `expertise` section if the CV explicitly shows percentage bars, star ratings, or numbered proficiency levels. Map to 0–100: - 5/5 stars → 100, 4/5 → 80, 3/5 → 60, 2/5 → 40, 1/5 → 20 - Percentages: use directly. ```jsonc { "id": "sec_000000X", "type": "expertise", "title": "Industry Expertise", "visible": true, "data": { "items": [ { "id": "xp_0000001", "label": "Machine Learning", "level": 80 } ] } } ``` ## Stage 4 — Output 1. **Write `src/data/default-resume.json`** in the project root (default behaviour). This is the file the app bundles as the seed shown on first visit and on Reset — no editor login required to see the result. - If `--output ` was specified, write to that path instead. 2. **Print the full JSON** to the conversation so the user can review it. 3. **Summarise what was imported** in a short table: | Field | Value | |------------------|----------------------------------------------------------| | Name | … | | Sections found | experience (N jobs), education (N), skills (N groups), … | | Dates normalised | N | | Fields to review | N (marked with `_review_*` keys) | 4. **Next steps** — tell the user exactly these two steps to preview the resume without logging in: Open `http://localhost:5173` in the browser, open DevTools console (F12 / Cmd+Option+I), paste this one line, and press Enter: ```js localStorage.removeItem("qwik-resume-editor:v2"); location.reload(); ``` This clears any cached resume from localStorage so the app falls back to the newly written `default-resume.json`. The page reloads and shows the imported resume on the public home page — no password needed. ## Error handling | Problem | Action | |-----------------------------|-------------------------------------------------------------------------------------------------------------------------------| | File not found | Report path, ask user to confirm | | PDF is scanned / image-only | Warn "no text layer detected"; suggest installing markitdown with OCR support (`pip install markitdown[pdf]`) or manual entry | | Date cannot be parsed | Use `""` and add `"_review_date": "original: "` | | Section heading ambiguous | Pick the closest match; add `"_review_section": "guessed from: "` | | markitdown exits with error | Fall back to LLM extraction; report the error message | | Contact type unclear | Default to `"string"` type; note it in the summary | ## Quality checks before output - [ ] All `id` fields are unique within the document - [ ] All dates are `"YYYY-MM"` or `""` — no other formats - [ ] `type` fields match the allowed `SectionType` enum exactly - [ ] `level` fields in `languages` use the exact enum values: `Basic | Elementary | Intermediate | Conversational | Proficient | Advanced | Native` - [ ] `level` in `expertise` items is a number 0–100 - [ ] `description` fields contain HTML, not plain text with hyphens - [ ] No section type appears more than once in `sections` - [ ] `version` is exactly `"2.0.0"` - [ ] `theme.paletteId` is one of: `orange-navy | navy-black | green-charcoal | purple-slate | magenta-graphite | mono` - [ ] JSON is valid (no trailing commas, no `//` comments in the actual output)