Build Resume

Interactive workflow: analyze a job description against the user's base resume (resume.md), identify strongest matches, fill gaps through conversation, and generate an ATS-optimized 2-page Word document.

Prerequisites

• resume.md in working directory (base resume — additive, never remove content)
• generate_resume.py, import_resume.py in working directory
• python-docx installed (pip3 install python-docx)
• For PDF import: pdfplumber installed (pip3 install pdfplumber)
• For Step 8 (Apply): playwright installed (pip3 install playwright), apply_to_role.py in working directory, Google Chrome installed

Importing a Resume

If resume.md does not exist and the user provides a resume file (.docx, .pdf, .txt), run the import flow before the main workflow:

1. Extract text from the file:

   python3 import_resume.py <resume_file>

2. Parse the extracted text into structured sections: name, contact info, overview/summary, experience (title, company, location, dates, description, bullets), education, and skills.
3. Write resume.md in this format:

   # Full Name
   
   phone | email | linkedin
   
   ## Overview
   Summary paragraph...
   
   ## Experience
   
   ### Job Title
   **Company | Location**
   *Dates*
   Role description...
   - Bullet 1
   - Bullet 2
   
   ## Education
   Degree — School
   
   ## Skills
   - Skill 1
   - Skill 2

4. Present resume.md to the user for review. Ask:
   • "Does this look complete? Anything missing or incorrect?"
   • "Any achievements, metrics, or skills to add?"
5. Apply corrections, then proceed with the main workflow.

Workflow

digraph {
  rankdir=TB; node [shape=box];
  fetch [label="1. Fetch & parse JD"];
  analyze [label="2. Analyze resume vs JD"];
  ask [label="3. Ask questions\n(1-2 rounds max)"];
  update [label="4. Update resume.md"];
  tailor [label="5. Tailor summary,\nbullets, & skills"];
  review [label="6. Present selection\nsummary for approval"];
  gen [label="7. Generate .docx"];
  apply [label="8. Apply to role\n(optional)", style="dashed"];
  fetch -> analyze -> ask -> update -> tailor -> review -> gen;
  gen -> apply [style=dashed, label="if user\nwants to apply"];
  ask -> ask [label="follow-up if needed"];
}

Steps

1. Fetch Job Description

• Try WebFetch on the URL first.
• Fallback chain (many job boards block automated fetching):
  1. WebFetch the URL directly
  2. If blocked/fails: WebSearch for the exact job title + company + "job description requirements" to extract JD details from search result snippets
  3. If still insufficient: ask the user to paste the full JD text
• Extract: job title, company, location, hard requirements, preferred qualifications, responsibilities, keywords.
• Distinguish hard requirements from nice-to-haves.

2. Analyze Resume Match

Read resume.md. For each JD requirement, classify resume coverage:

Category	Meaning
Covered	Strong bullet with quantified results directly maps to requirement
Partial	Related experience exists but weak fit or missing metrics
Gap	No relevant experience in resume.md

Select top 3-5 strongest matches. Prioritize:

1. Quantified achievements directly mapping to JD requirements
2. Leadership/scope matching the role level
3. Recent experience (last 10-15 years)
4. Unique differentiators

3. Ask Questions (batch efficiently, 1-2 rounds max)

Present your analysis first, then ask ALL questions in one round, grouped:

Gaps — "The JD requires X. Do you have experience with this?"

• Only ask where the candidate might plausibly have experience
• Explain why it matters for the role

Clarifications — "Your bullet about X — should this rank higher for Y requirement?"

• Borderline items where ranking could go either way

Deepening — "For [bullet], can you share [specific metric/dollar amount/percentage]?"

• When a number would strengthen a selected or likely-selected bullet

Additional skills — "Do you have experience with [JD-mentioned tools/technologies/certs] not in your resume?"

• Cover tools, frameworks, certifications, domain expertise from JD

If the first round surfaces significant new info, do ONE follow-up. Never more than 2 rounds.

4. Update Base Resume

• Add new bullets to the appropriate role section in resume.md
• Add new skills to the Skills section
• resume.md is additive only — never remove existing content
• These additions persist for all future resume builds

5. Tailor Content

Summary: Rewrite targeting THIS role. Mirror JD language. Hit top 3-4 keywords. 3-4 sentences max.

Bullet selection (within 10-15 year range):

• Current role: 4-6 bullets
• Previous 1-2 roles: 2-4 bullets each
• Older roles in range: 1-2 bullets or description only
• Beyond 15 years: title/company/dates under "Additional Experience"

Skills: Reorder with JD-matched skills first. Add user-confirmed skills. Remove irrelevant skills from tailored output only (keep in resume.md).

6. Present Selection Summary

Before generating, show:

• The 3-5 strongest bullets and which JD requirement each addresses
• Tailored summary text
• Ordered skills list
• Notable omissions and why they were cut
• Get user approval or adjustments

7. Generate Word Document

Primary method — if generate_resume.py exists:

python3 generate_resume.py /tmp/resume_input.json resume_[company]_[role].docx

• Company and role: lowercase, hyphens for spaces
• Clean up /tmp/resume_input.json after

Fallback — if generate_resume.py is not found, generate directly using docx-js (preferred) or python-docx:

Using docx-js (npm install docx):

• Follow the docx SKILL.md conventions for creating new documents
• Use US Letter page size (12240 x 15840 DXA) with 0.5" margins for maximum content space
• Use Arial font throughout: name 18pt bold (color #2E75B6), section headers 14pt bold with bottom border, job titles 11pt bold, body 10pt
• Structure: Name (centered) > Contact (centered) > Professional Summary > Professional Experience > Education > Core Competencies
• Use LevelFormat.BULLET for bullet points (never unicode bullets)
• Use TabStopType.RIGHT with TabStopPosition.MAX for right-aligned dates on job title lines
• Validate with python scripts/office/validate.py if available

Using python-docx (pip3 install python-docx):

• Create a clean, ATS-friendly single-column layout
• Use Pt(10) body text, Pt(14) section headers, Pt(18) name
• Set narrow margins: Inches(0.5) all sides for 2-page fit
• Add horizontal rules between sections using paragraph borders

8. Apply to Role (Optional)

After generating the resume, ask the user if they want to apply directly. If yes:

Setup (first time only):

pip3 install playwright

Automation loop — screenshot-driven:

1. Launch Chrome to the job URL:

   python3 apply_to_role.py launch "<job_url>"

2. Read the screenshot (/tmp/apply_screenshot.png) to understand the page state.
3. If login is required, tell the user: "Please log in to [site] in the Chrome window, then let me know when you're ready." After they confirm, take a new screenshot:

   python3 apply_to_role.py screenshot

4. Click the apply button:

   python3 apply_to_role.py click-apply

5. Read the screenshot to see the application form.
6. Upload the resume:

   python3 apply_to_role.py upload resume_[company]_[role].docx

7. Fill known fields from resume.md (name, email, phone, LinkedIn):

   python3 apply_to_role.py fill "<selector>" "<value>"

8. For dropdowns:

   python3 apply_to_role.py select "<selector>" "<option text>"

9. After each action, read the screenshot to verify and decide the next step.
10. For fields you can't determine (salary expectations, custom questions, etc.), ask the user.
11. If the page has a "Next" or multi-step flow, use click "text=Next" and repeat.
12. STOP before clicking Submit — always ask the user to review and confirm.
13. When done:

    python3 apply_to_role.py close

Other useful commands:

• text — extract page text (faster than screenshot for reading form labels)
• scroll down / scroll up — see more of the page
• pages / switch <index> — manage tabs if the site opens new ones
• url — check current page URL

Key rules:

• Always read the screenshot after every action — this is your feedback loop
• Never click Submit/Confirm without explicit user approval
• If automation fails for any action, tell the user to do it manually in the Chrome window, then screenshot to continue
• Chrome stays open between commands — the user can interact directly at any time

Job Tracker Integration (Optional)

When running repeated resume builds (especially via scheduled tasks), maintain a job application tracker spreadsheet (job_application_tracker.xlsx) as the single source of truth.

Tracker Schema

Column	Description
Date Found	Date the role was surfaced
Company	Company name
Job Title	Full job title
Location	City, State or Remote
Salary (if listed)	Salary range or blank
Job Posting URL	Must be a clickable hyperlink (see below)
Status	New, In Progress, Applied, Didn't Apply, Rejected, Interviewing, Offer
Reason (if Didn't Apply)	Why the user skipped this role
Resume File	Filename of the generated .docx
Date Applied	User fills this in
Last Contacted	Date of last interaction
Notes	Brief match summary

Writing Clickable URLs

When writing URLs to the tracker with openpyxl, always set both the value and the hyperlink:

cell = ws.cell(row=row, column=6, value=url)
cell.hyperlink = url
cell.font = Font(name="Arial", size=10, color="0563C1", underline="single")

Deduplication Rules

• Never suggest the same job posting URL twice (exact match against tracker)
• Never suggest a company that appeared in the tracker within the last 30 days
• If a company appears but it's been >30 days, a new role at that company is OK

Batch Search Mode (Scheduled/Multi-Role)

An alternate entry point for automated or scheduled runs that searches for multiple roles and lets the user pick which ones to build resumes for.

Workflow

1. Load tracker — read the job tracker spreadsheet to check for duplicates
2. Search — use WebSearch across multiple job titles, locations, and boards
3. Filter — enforce posting age (e.g., last 10 days), salary floor (e.g., $300K+), and deduplication against the tracker
4. Present — show the top N roles to the user with title, company, location, salary, and a 1-3 sentence summary of the role AND why it matches their background. Let user multi-select.
5. Log all — add every surfaced role to the tracker (selected = "New", skipped = "Didn't Apply")
6. Build — for each selected role, run the full Steps 1-7 workflow above (including Q&A — never skip it)
7. Update tracker — fill in the Resume File column and update status for completed builds

Search Strategy

• Search job boards: LinkedIn, Indeed, Greenhouse, Lever, Glassdoor, company career pages
• Prioritize roles posted within the last 48 hours
• Focus on roles matching the user's strongest differentiators (check resume.md for themes)
• Exclude roles clearly outside the user's domain unless the user has indicated interest

Scheduling

This mode can be configured as a scheduled task (e.g., 3x daily) using the schedule skill. The scheduled task prompt should include all search criteria, filtering rules, file paths, and the full build-resume workflow instructions so it runs autonomously.

ATS Content Rules

• Mirror exact JD phrases — if JD says "cross-functional collaboration," use that, not "working across teams"
• Include acronym + spelled-out — "Artificial Intelligence (AI)"
• Front-load keywords — most important keyword in first few words of each bullet
• Quantify everything — "$X revenue", "Y% growth", "Z team members"
• Use standard section names — Professional Summary, Professional Experience, Education, Core Competencies

Common Mistakes

Mistake	Fix
Picking impressive bullets over relevant ones	Score against JD requirements, not general impressiveness
Too many Q&A rounds	Batch all questions, max 2 rounds
Generic summary	Rewrite per role with JD keywords
Forgetting to update resume.md	Always persist new info before generating
Ignoring preferred qualifications	Still critical for ATS keyword matching
Assuming WebFetch will work on job boards	Most job sites block it — use the fallback chain
Writing plain-text URLs in the tracker	Always set cell.hyperlink for clickable links
Skipping Q&A in batch/scheduled mode	ALWAYS ask the user enhancement questions, even in batch mode