akbardevop/ai-job-agent/job-cv

Job CV

A skill (not an API call) for producing a tailored, ATS-friendly PDF resume for one specific posting. Claude does the rewriting. A small Node script renders the approved markdown to PDF via headless Chromium (Playwright). No LaTeX, no pandoc, no extra deps.

This pairs with /job-outreach (the cold-email skill): a tailored CV in hand makes the outreach concrete. It also feeds /job-apply — most ATS forms accept a custom resume upload.

Repo location

$AI_JOB_AGENT_ROOT → ~/.claude/skills/ai-job-agent/ → REPO_PATH marker file → ~/ai-job-agent/.

Prerequisites

Before drafting anything:

1. A base CV exists. Look in this order:

• $AI_JOB_AGENT_ROOT/cv.md (preferred)
• $AI_JOB_AGENT_ROOT/config/cv.md
• Extract a CV-shaped markdown from $AI_JOB_AGENT_ROOT/config/candidate-profile.md (sections: identity, education, experience, projects, skills) and write it to $AI_JOB_AGENT_ROOT/cv.md so future runs reuse it. Confirm with the user before writing.

If none of the above is reachable: tell the user to run /job-setup first and stop.

1. Chromium is reachable for Playwright. If npm test or a previous run errored with Executable doesn't exist, tell the user:

> `` > npx playwright install chromium > `` > then retry. Stop until set up.

1. output/ exists at $AI_JOB_AGENT_ROOT/output/. If not, create it (mkdir -p).

Workflow

Step 1 — Understand the target

Parse $ARGUMENTS. It might be:

• A direct job URL (LinkedIn, Greenhouse, Lever, Ashby, Jobvite, careers page)
• Pasted JD text in the user's message
• A company name + role ("backend intern at Stripe" — fetch the listing yourself)

If only a company is given and no URL, search for the listing first, then confirm the exact posting with the user before continuing.

Step 2 — Read the JD

• URL → use WebFetch to pull the page. If the result is a JS-only shell with no JD text, fall back to WebSearch for the same role at the same company and try the second hit.
• Pasted text → use as-is.
• Extract: company, role title, location / work-auth notes, 3-7 must-have skills (the ones most repeated or in the title), 3-5 nice-to-haves, tone (research-y? startup-y? enterprise?).

If you cannot find the JD body after two reasonable fetches, ask the user to paste it. Don't guess.

Step 3 — Read the base CV

Read the resolved base CV markdown. Note every concrete bullet, project, and skill. This is the bounding box — Step 4 may reorder, condense, or rephrase any of these, but cannot invent new ones.

Step 4 — Tailor (the actual job of this skill)

Produce a new markdown CV that:

1. Reorders sections and bullets so the JD's must-haves surface in the top half of the first page.
2. Rephrases existing bullets to use the JD's vocabulary (e.g., the CV says "computer vision pipeline"; the JD says "perception stack" — rewrite to match, but only if the underlying work is the same thing).
3. Drops sections / bullets that don't connect to the role. (E.g., a Quran subtitle editor side project is not landing on a substation engineering CV.)
4. Trims to one page when realistic. Two pages is fine if Step 3 has 5+ years of experience or many publications. Never zero-pad.
5. Preserves everything truthful: dates, titles, GPAs, links, work-auth phrasing.

Hard constraint — never fabricate:

• If the JD requires a skill the candidate does not have on the base CV, flag it to the user inline (e.g., "JD requires Rust; not on your CV. Should I (a) leave it off, (b) add a one-line 'Familiar with Rust syntax via X' if true, or (c) skip this posting?").
• Do not invent quantities, accomplishments, employers, or dates.
• Do not "round up" timelines. "3 months" stays "3 months."

Step 5 — Present the tailored CV for approval

Render exactly this structure in chat:

**Tailored CV draft** · for: <Role> at <Company> · base: cv.md (<line-count> lines) → tailored: <new-line-count> lines
 

<Name>

…full tailored markdown…

 
Changes vs base:
- Reordered: <section A> before <section B> (matches JD priority on <skill>)
- Rephrased: "<old bullet>" → "<new bullet>" (uses JD term <…>)
- Dropped: <section / bullet> (not relevant to <role>)
- Flagged gaps: <skill from JD not on CV> — what should we do?
 
Approve, edit, or cancel? (approve / edit / cancel)

Step 6 — Handle the response

• approve → go to Step 7.
• edit → ask "what should I change?" and regenerate Step 4 + 5. Loop until approve or cancel.
• cancel → do nothing, don't write any file. Tell the user "nothing rendered, nothing logged."

Step 7 — Render the PDF

1. Save the approved markdown to:

`` $AI_JOB_AGENT_ROOT/output/cv-<company-slug>-<YYYY-MM-DD>.md ``

Slugify the company name: lowercase, alphanumerics + dashes only.

1. Build the JSON payload for the renderer:

``json { "cv": "<approved markdown>", "outputPath": "<repo>/output/cv-<company-slug>-<YYYY-MM-DD>.pdf", "title": "Resume — <Candidate Name> — <Role> @ <Company>", "format": "letter" } ``

Default format is letter (US). Use a4 if the role is EU-based.

1. Render:

``bash echo "$PAYLOAD" | node "$AI_JOB_AGENT_ROOT/scripts/generate-tailored-cv.mjs" ``

Parse the JSON output. Expect ok: true, outputPath, bytes, pages. If ok: false with code: 3 and the error mentions "Executable doesn't exist", tell the user to run npx playwright install chromium and stop.

Step 8 — Log

Append a notes-column entry on the matching row of $AI_JOB_AGENT_ROOT/application-tracker.csv. If no row exists yet for this <company> + <role>, just store the path — /job-apply will create the tracker row when the user actually applies.

Notes-column format (semicolon-separated, append, don't replace):

tailored-cv: output/cv-<company-slug>-<YYYY-MM-DD>.pdf (<pages>p, <bytes>B)

Step 9 — Confirm and suggest next

Print:

Then nudge: "Want me to (a) draft a cold email to a hiring manager at this company via /job-outreach, (b) submit the application via /job-apply, or (c) tailor for the next posting?" — don't auto-do, wait.

Rules

• Never invent experience. Reorder, rephrase, drop. Never add. Flag JD-skill gaps explicitly.
• One page when realistic. Don't pad. Don't shrink the font below 10.5pt to fit. If it's two pages, that's fine.
• One column, serif body, no icons, no multicol. ATS parsers choke on fancy layouts. The default CSS in the renderer is already ATS-tested — don't override css unless the user asks.
• Truth-preserving rewrites only. "Computer vision pipeline" → "perception stack" is fine if the work is the same. "Built a CV pipeline" → "Led a perception team" is not.
• Never auto-send the PDF anywhere. This skill produces a file. The user decides what to do with it.
• Never overwrite the base `cv.md` with the tailored version. Tailored output always goes to output/.
• One PDF per posting per day. If output/cv-<slug>-<today>.pdf exists, append -v2, -v3, etc. — don't silently overwrite.

Related skills

• /job-setup — onboarding; writes the base cv.md if absent.
• /job-apply — submits the application via the right ATS filler. Pass it the tailored PDF path when the form has a resume-upload field.
• /job-outreach — companion cold-email skill; works great with a freshly-tailored CV in hand.
• /tailor-resume (community skill) — alternative tailoring flow that does not render PDF; use this skill if you want both rewrite + PDF in one shot.