Apply

Apply

Trigger

Keyword: apply

The user says apply (or variants: "apply to N jobs", "postulate", "search jobs") and the full search and application flow is triggered.

Pre-flight (applies to ALL applications: LinkedIn Easy Apply AND direct career pages)

• Verify active LinkedIn session. If session closed → open browser with wrapper (see AGENTS.md "Browser session"): node .agents/skills/browser-automation/scripts/browser.js open <url> --headed (Gold Rule 5) → notify user → wait for confirmation
• Browser: always use node .agents/skills/browser-automation/scripts/browser.js for open/close/goto. See AGENTS.md "Browser session" and "Parallel execution" for details. Never call agent-browser open directly, never open Chrome directly
• Parallel execution: if running alongside other flows (e.g: news or targets), pass --session apply-1 to linkedin-easy-apply.js and all browser.js exec commands. agent-browser sessions are created implicitly by the flag — no attach step needed. When done, just stop using the session (never close-all — that kills the browser for other agents too)
• Load active preferences (see memory skill):

  bashCopy code
  node scripts/db.js "SELECT category, key, value, confidence, source FROM preferences WHERE user_id = <user_id> AND status = 'active' ORDER BY category, key"
  

• Load strategy (see AGENTS.md "Strategy levels"):

  bashCopy code
  node scripts/db.js "SELECT data->'strategy' AS strategy FROM users WHERE id = <user_id>"
  

  Respect: apply_batch_size (max jobs per session), match_threshold (must_only / must_strong / must_strong_nice), relax_must_haves (loosen Must-have filtering). If apply_batch_size = 0, don't auto-apply, only present matches for manual approval
• Read profile and existing applications via db CLI:

  bashCopy code
  node scripts/db.js "SELECT data->'profile' AS profile, data->'job_preferences' AS prefs, data->'personal_info' AS personal FROM users WHERE id = <user_id>"
  node scripts/db.js "SELECT url FROM applications WHERE user_id = <user_id>"
  

• DB is the single source of truth for ALL form fields. Before filling ANY form (LinkedIn, Lever, Greenhouse, Workday, SuccessFactors, custom sites), the agent must have the profile data loaded in context. Never invent, guess, or fabricate any value. If a required field is not in the DB, STOP, ask the user, save the answer to DB, then continue. This is Gold Rule 5c.
• Captcha policy: NEVER attempt to solve captchas programmatically. This is Gold Rule 5b. When a captcha appears (hCaptcha, reCAPTCHA, image challenge, drag-and-drop, etc.), the agent must: (1) ensure browser is headed, (2) notify the user and wait, (3) continue only after user confirms. Never retry in a loop. Never attempt to click captcha elements, solve challenges, or bypass them.

Flow

1. Search jobs

Preferred: use the automation script (see AGENTS.md "Scripts de automatizacion"):

bashCopy code
# Dry-run first to see what's available
node scripts/linkedin-easy-apply.js --dry-run --max 15

# Apply to top N jobs
node scripts/linkedin-easy-apply.js --max 10

# Run in a specific browser session (for parallel execution with other agents)
# Sessions are implicit in agent-browser — just pass --session apply-1 to every command
# Then run the script with --session:
node scripts/linkedin-easy-apply.js --max 10 --session apply-1

The script handles: search with Easy Apply filter, title filtering (only applies to tech roles — a job must match a tech keyword and not hit the blocklist of non-tech roles like attorney/sales/tutor/support; extendable via job_preferences.avoid in DB), form filling with standard answers, radio/combobox/checkbox handling, DB registration, and captcha detection (stops on captcha). The --session flag allows running in parallel with other agents by using an attached session instead of the default one.

Manual fallback (when script fails or forms have complex open-ended questions):

Search LinkedIn Jobs with filters:

• Keywords derived from profile (primary role, seniority, AI-related terms)
• Location: Worldwide or remote
• Work type: Remote
• Experience level: per profile (mid-senior, director)
• Easy Apply: yes (filter f_AL=true)
• Sort: Date posted (most recent first)

Paginate until collecting 20-30 candidates.

2. Filter by Must-haves

For each job, verify against profile Must-haves from users.data.job_preferences. Discard if:

• Doesn't match any Must-have (the filter is dynamic, based on the user's Must-weighted preferences, not hardcoded)
• Requires visa/location the user doesn't have (e.g: US-only, EU-only)
• Not in the user's field/sector (per users.data.profile.sector and job_preferences)
• Already applied (check against DB)

Keep 10-15 matching positions.

2.5. Warm Sourcing & Referral Pre-Check (Strategy #1 & #4)

Gate: only run this step if referrals is in strategy.sources_active. If not, skip to Step 3.

For each selected position, before submitting a cold application:

1. Run warm sourcing discovery:

   bashCopy code
   node scripts/linkedin-warm-sourcing.js --company "<Company>" --role "<Role>" --json
   

2. If an internal contact, alumni, or ex-colleague is found:
   • Prioritize Strategy #1 (Internal Referral) over cold apply.
   • Stage a personalized referral request draft in messages table (following Gold Rule 7 & user style profile).
   • Register card in pipeline as discovered.
3. If NO internal contact exists:
   • If strategy.cold_outreach = false → skip recruiter outreach, proceed to Step 3 (cold apply only).
   • If strategy.cold_outreach = true → extract Recruiter / Hiring Manager info for the position, stage a recruiter outreach DM draft (Strategy #4 Multi-channel combo).
   • Perform ATS micro-alignment (tailor CV keywords to JD if applying directly/email) and proceed to Step 3.

3. Apply via Easy Apply

For each selected job:

1. Navigate to the job URL
2. Click "Easy Apply"
3. Advance through form steps:
   • Contact info: pre-filled by LinkedIn, verify
   • Resume: already loaded in LinkedIn profile, verify
   • Additional questions: answer based on profile and preferences
     • Years of experience: use real value from profile
     • Salary expectations: use profile range from DB (job_preferences.salary.value: min/max/currency). Convert to local currency if the form requires it. Never invent a salary number.
     • Location: use personal_info.city + personal_info.country from DB
     • Work authorization: answer honestly
     • Availability: use job_preferences.availability.value from DB
   • For any field not in the DB: STOP, ask the user, save answer to DB, then fill. Gold Rule 5c.
4. Review → Submit
5. If a captcha appears at any point: STOP, ensure browser is headed, notify user, wait. Gold Rule 5b. Never attempt to solve it.
6. Verify "Application submitted" on screen
7. Register in DB via db CLI:

bashCopy code
node scripts/db.js "INSERT INTO applications (user_id, platform, company, role, url, status, data) VALUES (<user_id>, 'linkedin', '<company>', '<role>', '<url>', 'applied', '<json>'::jsonb)" --write

data should include: match reason, method (easy_apply), location, questions_answered count.

4. Anti-ban

• Wait 2-3 seconds between actions (don't spam clicks)
• Don't apply to more than 15 jobs per session
• If a captcha or block appears: STOP IMMEDIATELY. Do NOT attempt to solve it. Do NOT retry in a loop. Ensure browser is headed, notify the user, and wait for them to solve it. This is Gold Rule 5b. The agent fills the entire form, triggers submit, and when the captcha appears, it stops and asks the user. Period.
• Vary navigation order (don't go sequentially through the results list)

5. Summary

Upon completion, present table with:

• Company, role, URL
• Total applied
• Total skipped with reason

6. Direct applications (non-LinkedIn: Lever, Greenhouse, Workday, SuccessFactors, custom sites)

When applying directly to a company career page (not via LinkedIn Easy Apply), the same pre-flight rules apply. The flow is:

1. Load profile data from DB first (pre-flight checklist). Have all values in context before opening the form.
2. Navigate to the application URL.
3. Fill ALL form fields using ONLY data from the DB:
   • Name: profile.full_name or personal_info
   • Email: profile.email
   • Phone: personal_info.phone
   • Address: personal_info.address, personal_info.city, personal_info.state, personal_info.postal_code, personal_info.country
   • Salary: job_preferences.salary.value (min/max/currency). Convert to local currency if the form requires it.
   • LinkedIn URL: profile.linkedin_profile
   • GitHub URL: profile.github
   • CV: profile.cv_path or personal_info.cv_pdf_path
   • Work experience: profile.experience[]
   • Education: profile.education[]
   • Skills: profile.skills[] or profile.tech_stack
   • Languages: profile.languages[]
4. If a field is required but NOT in the DB: STOP, ask the user, save to DB, then fill. Gold Rule 5c. Never invent.
5. Upload CV when prompted.
6. Submit the form.
7. If a captcha appears: STOP, ensure headed, notify user, wait. Gold Rule 5b. Never solve programmatically.
8. Verify submission confirmation on screen.
9. Register in DB via db CLI (same INSERT as LinkedIn flow, with platform = the ATS detected: 'lever', 'greenhouse', 'workday', etc.).

6.1. Known ATS failure modes (validated 2026-09-19)

• Lever + hCaptcha rejects the automation browser fingerprint. Even a correctly human-solved challenge returns "verification failed" — the problem is the browser fingerprint, not the captcha answer. Do NOT retry in a loop. Workarounds, in order: (a) search the same req on LinkedIn Jobs — most Lever postings are cross-posted with Easy Apply; (b) check the company careers page for a general "Submit CV" / talent pool form (often has no captcha); (c) apply from the user's normal browser manually.
• Captcha tokens expire in minutes. Solve challenges immediately after triggering them, and one form at a time — solving two forms' captchas back-to-back kills the second token.
• Failed submits clear file inputs. After any submit error, re-attach the CV before retrying — the <input type=file> value does not survive validation failures.
• Forms asking for "real numbers / traction" (e.g. "users you got, with figures"): fetch real metrics via public APIs instead of guessing — npm https://api.npmjs.org/downloads/point/last-month/<package>, GitHub https://api.github.com/users/<user>/repos (stars, forks). Never invent figures (Gold Rule 5c applies to metrics too).
• Human-only deliverables (demo video, recorded answers, portfolio review calls): fill every other field first, keep the form open, register the application in DB with status='pending_user_action' and data.blocker describing the exact artifact needed, then ask the user for it at the end of the round (Gold Rule 5d).

Dependencies

• Depends on onboarding (DB to register)
• Depends on profile (Must-haves to filter)
• Consumed by daily

Easy Apply form answers (DB keys)

All personal data lives in the DB, never in scripts or docs. The agent and scripts read form answers from:

Common Easy Apply question types and where to get the answers:

• Years of experience with [tech]: users.data.form_answers.<tech>_experience
• Language level: users.data.form_answers.english_level / spanish_level
• Current location: users.data.form_answers.location
• Current company: users.data.form_answers.current_company
• LinkedIn URL: users.data.form_answers.linkedin_url
• Salary expectation: users.data.form_answers.salary_usd / salary_cop / salary_usd_max
• Availability: users.data.form_answers.notice_period / availability_date
• Consent/privacy: always accept
• Diversity/accessibility: users.data.form_answers.diversity_* (accessibility, gender, ethnicity)
• Disability: users.data.form_answers.disability
• GenAI tools experience: users.data.form_answers.genai_tools
• AWS experience: users.data.form_answers.aws_experience
• English comfort (open text): users.data.form_answers.english_comfort

If a key doesn't exist in form_answers: the script skips the field (doesn't invent it). The agent must stop, ask the user, save the answer to DB (jsonb_set on users.data.form_answers), then continue. Gold Rule 5c.

Script reference

scripts/linkedin-easy-apply.js -- Apply via Easy Apply

Searches jobs with Easy Apply filter, clicks, fills forms with standard answers, submits, registers in DB.

bashCopy code
# Apply to the first 10 jobs (default)
node scripts/linkedin-easy-apply.js

# Keywords custom + limit
node scripts/linkedin-easy-apply.js --keywords '"<Role>" OR "<Skill>"' --max 5

# List only, don't apply
node scripts/linkedin-easy-apply.js --dry-run

# Output JSON
node scripts/linkedin-easy-apply.js --json

Flags: --keywords <q> (default: derived from DB profile.title + profile.skills), --location <loc> (default: from DB job_preferences.location), --max <n> (default 10), --dry-run, --json, --session <name> (for parallel execution) Auto-fill: all values are read from users.data.form_answers (DB). The script fills: years of experience per tech, language level, location, current company, LinkedIn URL, salary, availability, GenAI tools, AWS, etc. Radios: Yes for skills, No for disability/sponsorship (configurable values in DB). Comboboxes: English/Spanish level, seniority (from DB). Captcha: detects and stops with exit 1 + message. Never attempts to solve. DB: registers each application with platform='linkedin', status='applied'.

scripts/linkedin-search.js -- Search posts for job openings

Searches LinkedIn posts, extracts author + vanity + email + content preview. Filters by relevance (AI/ML keywords) and dedupes.

bashCopy code
# Basic search (human-readable output)
node scripts/linkedin-search.js '"<Role>" "hiring" <Region>'

# Search with more scrolls and JSON output (to pipe to other scripts)
node scripts/linkedin-search.js '"<Role>" "<City>" "hiring"' --scroll 3 --json

# Validated queries:
#   '"<Role>" "hiring" <Region>'               (most productive)
#   '"<Role>" "<City>" "hiring"'      (geo-specific)
#   '"ingeniero IA" "buscamos"'                  (Spanish)

Flags: --scroll <n> (default 2), --json (raw JSON output) Output JSON: [{author, vanity, email, content}, ...]