Polish

Polish — LinkedIn profile + CV optimization

Trigger

Keyword: polish (or variants: "mejorar mi linkedin", "pulir perfil", "alinear cv", "optimizar perfil")

Purpose

Takes the user's captured profile (users.data.profile) and job preferences (users.data.job_preferences) and uses them to optimize the two artifacts that recruiters see: the LinkedIn profile and the CV. This is an output flow, not an input flow — profile captures data, polish applies it externally.

Dependencies

• onboarding (DB, browser profile, LinkedIn session)
• profile (requires users.data.profile and users.data.job_preferences with Must/Strong/Nice weights)

Parallel execution

polish can run alongside other flows (e.g: apply, news, targets) by using an attached session:

bashCopy code
node .agents/skills/browser-automation/scripts/browser.js open <url>
node .agents/skills/browser-automation/scripts/browser.js exec eval '<code>' --session polish-1
node scripts/generate-cv.js --session polish-1
node .agents/skills/browser-automation/scripts/browser.js exec close --session polish-1

All browser commands and generate-cv.js accept --session. When done, close only your own session (exec close --session polish-1) or leave it — never close-all, which kills the browser for other agents. See AGENTS.md "Parallel execution".

Gate de validacion (pre-flight obligatorio)

Before executing any phase, verify that dependencies are satisfied. If any check fails, do not proceed — tell the user what is missing and how to resolve it:

bashCopy code
# 1. Verify onboarding completed: DB exists and has user
node scripts/db.js "SELECT id, name, email, data FROM users WHERE id = <user_id>"
# If no row → "Necesitas ejecutar `onboarding` primero. No hay DB configurada."

# 2. Verify profile exists with minimum data
node scripts/db.js "SELECT data->'profile' AS profile, data->'job_preferences' AS prefs FROM users WHERE id = <user_id>"
# If profile is null/empty → "Necesitas ejecutar `profile` primero. No hay perfil capturado."
# If job_preferences is null/empty → "Necesitas completar el cuestionario de `profile`. No hay preferencias declaradas."

# 3. Verify minimum fields within profile
# Required for Phase 1 (LinkedIn): profile.title, profile.experience[], profile.skills[]
# Required for Phase 2 (CV): profile.full_name, profile.email, profile.experience[], profile.education[]
# If any required field missing → "Tu perfil esta incompleto. Falta: <fields>. Ejecuta `profile` para completarlo."

# 4. Verify LinkedIn session is active
node .agents/skills/browser-automation/scripts/browser.js status   # lists active sessions
# If fails → "Necesitas iniciar sesion en LinkedIn. Ejecuta `onboarding` o abre el browser headed para login."

# 5. Verify linkedin_profile URL exists in DB
node scripts/db.js "SELECT data->'linkedin_profile' AS url FROM users WHERE id = <user_id>"
# If null → "No tengo tu URL de LinkedIn. Ejecuta `onboarding` para guardarla."

Only if all 5 checks pass, continue to Phase 1.

Phase 1 — LinkedIn profile optimization

1a. Audit (read-only)

1. Load from DB: profile, job_preferences, linkedin_profile, style_profile, strategy (for strategy_level)
2. Navigate to the user's LinkedIn profile: node .agents/skills/browser-automation/scripts/browser.js goto <linkedin_profile_url>
3. Take a snapshot to understand the current page structure: node .agents/skills/browser-automation/scripts/browser.js exec snapshot
4. Extract current state of each section using eval (adapt selectors to what you see in the snapshot):

   bashCopy code
   node .agents/skills/browser-automation/scripts/browser.js exec eval '(function(){
     // Adapt selectors based on current LinkedIn DOM.
     // LinkedIn changes their UI frequently, so read the snapshot first
     // and adjust these selectors as needed.
     var headline = document.querySelector("h1")?.textContent?.trim() || "";
     var about = document.querySelector("#about ~ * .display-text, #about + * .inline-show-more-text")?.textContent?.trim() || "";
     // Experience: iterate over section entries
     var expNodes = document.querySelectorAll("#experience ~ * .pvs-entity, [data-view-name*='experience'] .pvs-entity");
     var experience = Array.from(expNodes).map(function(n) {
       return {
         title: n.querySelector(".t-14 .t-bold span")?.textContent?.trim() || "",
         company: n.querySelector(".t-14:not(.t-bold) span")?.textContent?.trim() || "",
         description: n.querySelector(".t-14.t-normal.t-black--light span")?.textContent?.trim() || ""
       };
     });
     // Skills
     var skillNodes = document.querySelectorAll("#skills ~ * .pvs-entity, [data-view-name*='skill'] .pvs-entity");
     var skills = Array.from(skillNodes).map(function(n) {
       return n.querySelector(".t-14 .t-bold span")?.textContent?.trim() || "";
     }).filter(Boolean);
     return JSON.stringify({ headline: headline, about: about, experience: experience, skills: skills });
   })()'
   

   • The eval code above is a starting point. Always take a snapshot first and adapt selectors to the current DOM. LinkedIn changes their class names frequently.
   • Extract: headline, about, experience (each role: title, company, period, description), education, skills (list + top 3 pinned), featured, open to work (if active, which roles), languages, certifications
5. Save snapshot to DB: users.data.linkedin_snapshot
6. Gap analysis: compare current state vs objectives:
   • Does headline reflect target role + AI focus?
   • Does About have a clear pitch aligned to job_preferences.role_types and ai_focus?
   • Does Experience have quantified achievements or just task descriptions?
   • Do Skills include those from job_preferences.stack and AI-related skills?
   • Is Open to Work active with the correct roles (if strategy is active/aggressive)?
7. Present gap report to user with specific recommendations

1b. Apply improvements (with per-section approval)

For each section with gaps, draft all changes for that section and show them together to the user for approval:

1. Headline: draft 2-3 options aligned to profile.title + top skills + job_preferences.ai_focus. Example: "<Title> | <Top Skills> | <Work Mode>"
2. About: draft 3-4 paragraph summary positioning the user for target roles, mentioning AI focus if Must, ending with a soft CTA
3. Experience: for each role, rewrite descriptions as quantified achievements (format: "Action + Context + Result"). Use data from original CV (profile.experience[])
4. Skills: reorder to put the most target-aligned skills in top 3. Add missing skills from job_preferences.stack
5. Open to work: if strategy_level is active or aggressive, activate "Open to work" with roles from job_preferences.role_types and job_preferences.seniority

Per-section approval flow:

• Show all changes for the section (before → after for each field)
• User approves the entire section, rejects it, or requests edits
• If approved: navigate to the section's edit URL, apply changes via eval (see below)
• Save each applied change to users.data.linkedin_polish_log (audit trail with before/after)

LinkedIn edit URLs

LinkedIn uses direct URLs to edit each section:

• Headline: https://www.linkedin.com/in/<vanity>/edit/details/ → click pencil icon on headline
• About: https://www.linkedin.com/in/<vanity>/edit/details/ → click pencil icon on about
• Experience: https://www.linkedin.com/in/<vanity>/edit/details/experiences/
• Skills: https://www.linkedin.com/in/<vanity>/edit/details/skills/
• Open to work: https://www.linkedin.com/in/<vanity>/edit/details/recruiteroptin/

How to edit LinkedIn sections via eval

LinkedIn editors are contenteditable (tiptap/slate). The agent interacts with them via node .agents/skills/browser-automation/scripts/browser.js exec eval '<code>'. Always take a snapshot first to find the correct refs/selectors, then:

1. Click the edit button (pencil icon) via eval:

   bashCopy code
   node .agents/skills/browser-automation/scripts/browser.js exec eval 'document.querySelector("button[aria-label*=\"Edit\"]").click()'
   

2. Fill the input/contenteditable with the new text:

   bashCopy code
   # For text inputs (headline):
   node .agents/skills/browser-automation/scripts/browser.js exec eval '(function(){
     var input = document.querySelector("input[type=\"text\"]");
     input.value = "<new headline text>";
     input.dispatchEvent(new Event("input", {bubbles: true}));
     input.dispatchEvent(new Event("change", {bubbles: true}));
   })()'
   
   # For contenteditable (about, experience descriptions):
   node .agents/skills/browser-automation/scripts/browser.js exec eval '(function(){
     var editor = document.querySelector("[contenteditable=\"true\"]");
     editor.focus();
     editor.textContent = "<new text>";
     editor.dispatchEvent(new InputEvent("input", {bubbles: true, inputType: "insertText"}));
     editor.dispatchEvent(new Event("change", {bubbles: true}));
   })()'
   

3. Click Save via eval:

   bashCopy code
   node .agents/skills/browser-automation/scripts/browser.js exec eval 'document.querySelector("button[type=\"submit\"], button[aria-label*=\"Save\"]").click()'
   

These are starting points. Always take a snapshot after navigating to the edit page and adapt selectors to what you see. LinkedIn's DOM changes frequently. The agent's advantage over a hardcoded script is that it can adapt to the current DOM in real time.

Phase 2 — CV optimization

2a. Analyze current CV

1. Read current CV from profile.cv_path (PDF) or profile.cv_url
2. Extract structure: summary, experience, education, skills, projects
3. Compare vs LinkedIn snapshot (from Phase 1a) and vs job_preferences
4. Identify gaps:
   • Does the CV summary position for the target role?
   • Does experience use impact verbs and quantification?
   • Are key target stack skills missing?
   • Is there irrelevant experience that dilutes the message?
   • Is the format ATS-friendly (selectable text, no complex tables)?

2b. Draft improved CV

1. Generate CV in Markdown format (intermediate, reviewable):
   • Header: name, title, contact (email, phone, LinkedIn, GitHub, blog)
   • Summary: 2-3 lines aligned to target role + AI focus
   • Experience: each role with 3-5 bullets of quantified achievements
   • Skills: grouped by category (Languages, AI/ML, Cloud, Tools)
   • Education: degree, institution, year
   • Projects: 2-3 relevant projects with impact
   • Languages: with proficiency level
2. Show the drafted CV to the user for review (rendered, not raw Markdown)
3. Iterate if the user requests changes
4. Save the final CV to:
   • users.data.cv_markdown (Markdown content, for future iterations)
   • PDF file generated via browser headless

2c. PDF generation via browser headless

The PDF flow uses scripts/generate-cv.js:

1. Convert Markdown to HTML with clean CV CSS (ATS-friendly, single page if possible)
2. Write HTML to a temp file
3. Open browser headless: node .agents/skills/browser-automation/scripts/browser.js open file://<path> --headless
4. Export to PDF via agent-browser (exec pdf <path>)
5. Save PDF path to users.data.cv_path (updates existing path)
6. Close browser

The user never sees Markdown or HTML. They see only the final PDF. If they want adjustments, they tell the agent what to change and the agent regenerates.

2d. CV tailoring per application (future, not part of this flow)

The optimized base CV is generic to the target role. For specific applications, the apply or targets flow can do "light tailoring" of the base CV (reorder skills, adjust summary to mention the company). This is documented as a future extension, not implemented now.

Persistence in DB

New JSONB keys in users.data:

Existing keys that get updated:

Final report

Copy code
## Polish report

### LinkedIn profile
| Section | Status | Changes applied |
|---|---|---|
| Headline | Updated | "<old>" → "<new>" |
| About | Updated | Added AI focus paragraph + CTA |
| Experience (3 roles) | Updated | Rewrote 8 bullets as quantified achievements |
| Skills | Reordered | Pinned: <skill1>, <skill2>, <skill3> |
| Open to work | Activated | Roles: <role1>, <role2>, <role3> |

### CV
- Format: Markdown → PDF (via headless browser)
- Sections optimized: summary, experience, skills
- Saved to: users.data.cv_markdown + <pdf_path>

### Pending (need attention)
- [manual] LinkedIn "Featured" section: add 2-3 projects (requires manual curation)

Rules

• Per-section approval: show all changes for a section together. User approves the entire section or rejects it
• Gold Rule 5c: do not invent quantified achievements. If the CV has no metrics, draft the bullet without a number and mark it as [TODO: add metric] for the user to complete
• Gold Rule 9: the flow is generic. All data comes from DB. Examples in this SKILL use <placeholder> syntax
• Anti-LLM style (Gold Rule 7): the About and experience bullets must sound human, not generated. No em-dashes, no bullet lists in About (it is prose), no obvious JD keywords
• Resumable: each change is persisted individually. If interrupted, the next run continues from where it left off
• Non-destructive: always save the before in linkedin_polish_log. The user can revert
• Single user (repo owner)

Scripts

scripts/generate-cv.js

Converts cv_markdown from DB (or --markdown <path>) to a PDF via browser headless.

bashCopy code
node scripts/generate-cv.js [--output <path>] [--markdown <path>] [--session <name>]

If --markdown is not provided, reads users.data.cv_markdown from DB. If --output is not provided, saves to .browser-profile/cv-polished-<timestamp>.pdf. Updates users.data.cv_path in DB after generating.

This is the only script in the polish flow. LinkedIn profile audit and editing are done directly by the agent via node .agents/skills/browser-automation/scripts/browser.js exec eval and node .agents/skills/browser-automation/scripts/browser.js exec snapshot, which allows the agent to adapt to LinkedIn's DOM in real time rather than relying on hardcoded selectors.

Dependencies

• Depends on onboarding (DB, browser profile, LinkedIn session)
• Depends on profile (requires users.data.profile and users.data.job_preferences)
• apply and targets can consume cv_markdown and cv_path for future tailoring