Resume Builder skill for Codex

Done well means: an ATS extracts every section in the right order with correct dates and contact details, a recruiter scanning for six seconds sees role, scope and results, every claim is true and supplied by the user, and DOCX and PDF come from the same source.

Hard rules

• Never invent employers, titles, dates, degrees, skills or numbers. When a bullet needs a metric the user has not given, write [metric?] and ask. Tailoring means choosing and phrasing true material, not adding new claims.
• Resumes are personal data. Keep files local; do not paste contents into online tools or commit them to public repositories.
• Follow regional norms the user confirms: US/UK/Canada usually omit photo, birth date and marital status; some European markets expect a photo or more personal details.

Workflow

- [ ] 1. Collect source material: old resume, LinkedIn export, target job description
- [ ] 2. Normalise into resume.yaml (schema below); confirm facts with the user
- [ ] 3. Write or rewrite bullets (achievement formula); mark gaps with [metric?]
- [ ] 4. Tailor: map job-description keywords to true experience; order by relevance
- [ ] 5. Render DOCX (python-docx) and PDF (Typst or HTML + WeasyPrint)
- [ ] 6. Run the ATS parse check on both files
- [ ] 7. Deliver files + a short list of open questions

Source of truth

# resume.yaml - loosely follows the JSON Resume schema
basics:
  name: Ada Lovelace
  label: Senior Backend Engineer
  email: [email protected]
  phone: "+44 20 7946 0000"
  location: London, UK
  links: [linkedin.com/in/adalovelace, github.com/ada]
summary: >-
  Backend engineer with 9 years in payments; led the ledger rewrite that cut
  reconciliation time from days to minutes.
work:
  - company: Analytical Engines Ltd
    position: Senior Backend Engineer
    location: London
    start: 2021-03
    end: present
    highlights:
      - Led a 5-person team rebuilding the payments ledger in Go; reconciliation time fell from 2 days to 15 minutes.
      - Cut p99 API latency from 900 ms to 180 ms by replacing N+1 queries and adding read replicas.
education:
  - institution: University of London
    area: Mathematics
    studyType: BSc
    end: 2015
skills:
  - {name: Languages, keywords: [Go, Python, SQL]}
  - {name: Infrastructure, keywords: [PostgreSQL, Kafka, AWS, Terraform]}

Writing bullets

Formula: verb + what you did + scope + measurable result.

Weak	Strong
Responsible for the billing service	Owned the billing service (40k invoices/month); cut failed charges 35% by adding retry with idempotency keys
Worked on performance	Reduced p99 checkout latency from 1.2 s to 300 ms by caching price rules
Helped onboard new hires	Wrote the onboarding guide and pairing plan; new engineers shipped first PR in 3 days instead of 2 weeks

• 3-6 bullets for recent roles, 1-2 for older ones; drop roles older than ~15 years to a one-line “Earlier experience” entry.
• Length: one page under ~10 years of experience, two pages beyond. Never three.
• Tailoring: list the job description’s hard requirements, find the true evidence for each, move it up, and use the employer’s wording where it is accurate (ATS keyword match is literal: “PostgreSQL” vs “Postgres”, “CI/CD” vs “continuous delivery”).

ATS-safe layout

Do	Avoid
Single column, top-to-bottom reading order	Two-column layouts, sidebars
Real text, standard fonts (Calibri, Arial, Garamond, Source Sans)	Text in images, icons for phone/email
Headings named Experience, Education, Skills, Projects, Certifications	Creative headings (“My Journey”)
Contact details in the body at the top	Contact details only in the page header/footer
Dates as Mar 2021 - Present or 2021-03 - present, one style throughout	Years only on some roles; “Spring ‘21”
.docx when the portal asks for Word; text-based PDF otherwise	Scanned PDF, tables for layout, text boxes

Render DOCX with python-docx

from datetime import datetime
import yaml
from docx import Document
from docx.enum.text import WD_TAB_ALIGNMENT
from docx.shared import Cm, Pt

def fmt(d):
    if d in (None, "present"):
        return "Present"
    return datetime.strptime(str(d), "%Y-%m").strftime("%b %Y") if "-" in str(d) else str(d)

cv = yaml.safe_load(open("resume.yaml", encoding="utf-8"))
doc = Document()
sec = doc.sections[0]
sec.left_margin = sec.right_margin = Cm(2)
sec.top_margin = sec.bottom_margin = Cm(1.6)
normal = doc.styles["Normal"]
normal.font.name, normal.font.size = "Calibri", Pt(10.5)
normal.paragraph_format.space_after = Pt(2)
right_tab = sec.page_width - sec.left_margin - sec.right_margin

b = cv["basics"]
doc.add_heading(b["name"], level=0)
doc.add_paragraph(" | ".join([b["label"], b["location"], b["email"], b["phone"], *b["links"]]))
doc.add_heading("Summary", level=1)
doc.add_paragraph(cv["summary"])

doc.add_heading("Experience", level=1)
for job in cv["work"]:
    p = doc.add_paragraph()
    p.paragraph_format.tab_stops.add_tab_stop(right_tab, WD_TAB_ALIGNMENT.RIGHT)
    p.paragraph_format.space_before = Pt(6)
    p.add_run(f"{job['position']}, {job['company']}").bold = True
    p.add_run(f"\t{fmt(job['start'])} - {fmt(job['end'])}")
    for h in job["highlights"]:
        doc.add_paragraph(h, style="List Bullet")

doc.add_heading("Education", level=1)
for ed in cv["education"]:
    doc.add_paragraph(f"{ed['studyType']} {ed['area']}, {ed['institution']}, {ed['end']}")

doc.add_heading("Skills", level=1)
for s in cv["skills"]:
    p = doc.add_paragraph()
    p.add_run(f"{s['name']}: ").bold = True
    p.add_run(", ".join(s["keywords"]))

doc.core_properties.title = f"{b['name']} - Resume"
doc.core_properties.author = b["name"]
doc.save("Ada_Lovelace_Resume.docx")

Contact details sit in the body, the right-aligned dates use a tab stop (not a table), and headings use real Heading styles. Restyle Title and Heading 1 in the document’s styles (size, colour, a bottom border) rather than formatting runs one by one.

Render PDF

Generate the PDF from the same resume.yaml: a Typst template (yaml("resume.yaml") loads it directly; see $typst-documents) or Jinja2 HTML + WeasyPrint (see $pdf-generation). Converting the DOCX with LibreOffice also works (soffice --headless --convert-to pdf Ada_Lovelace_Resume.docx) and keeps both files identical. Embed fonts, keep one column, and set the PDF title metadata to the person’s name.

File names: Firstname_Lastname_Resume.pdf, not resume_final_v3.pdf.

ATS parse check

Extract text the way an ATS does and read it top to bottom:

pdftotext -layout Ada_Lovelace_Resume.pdf - | head -60
python3 -c "import docx,sys; print('\n'.join(p.text for p in docx.Document(sys.argv[1]).paragraphs))" Ada_Lovelace_Resume.docx

Pass criteria:

• Name, email and phone appear in the first lines of extracted text.
• Section headings appear as their own lines in the expected order.
• Each role’s title, company and dates are on adjacent lines, not split across columns.
• No (cid: codes, missing characters or merged words (font or ligature problems).
• Every hard requirement from the job description appears verbatim at least once, backed by a true bullet.
• pdffonts shows embedded fonts; page count is 1 or 2.

Delivery note

**Files:** Ada_Lovelace_Resume.docx, Ada_Lovelace_Resume.pdf (1 page, from resume.yaml)
**Tailored for:** Staff Engineer, Payments - matched 9/10 listed requirements
**ATS check:** text order correct in both files; contact details in body; fonts embedded
**Needs your input:** [metric?] on the 2019 migration bullet; confirm the Kafka certification date