Generate Application

generate-application

Generate a complete, targeted application package for a job posting. The output lives at applications/YYYY-MM-DD-[company]-[role-slug]/ in the project root and includes a deep analysis, a targeted CV data file, a rendered CV PDF, drafted answers to the actual application form questions, and a two-paragraph cover letter.

A job ad is not the application. The real application lives on a separate apply deep link (Greenhouse #app form, Lever /apply, Ashby /application, Workable apply page, …) with screening questions, essay prompts, and factual fields the ad never mentions. A package that skips these is incomplete — Step 2 is mandatory, not optional.

Step 1 — Fetch and Parse the Job Posting

Fetch the job posting using Jina Reader for clean markdown output:

curl -s -H "Authorization: Bearer $JINA_API_KEY" "https://r.jina.ai/[job-url]"

If JINA_API_KEY is not set, fall back to WebFetch with the raw URL.

Extract from the posting:

• Company name and role title
• Required and preferred skills/technologies
• Seniority signals (years experience, leadership expectations)
• Work arrangement (remote, hybrid, location)
• Mission/culture signals

Step 2 — Fetch the Application Form (Deep Link) and Extract Questions

Discover the apply deep link and pull the actual application questions:

bash .claude/skills/generate-application/scripts/fetch-application-form.sh "[job-url]"

Returns JSON: {url, ats, apply_url, source, needs_browser, questions, form_text}. The script uses the ATS's public API where one exists (Greenhouse ?questions=true, Workable form API, Ashby GraphQL posting API — all return labels, required flags, and dropdown options) and falls back to scraping the apply page.

Interpret the result:

• needs_browser: false — questions is authoritative; use it directly.
• needs_browser: true — the form wasn't captured (empty questions; form_text is at best a thin JD blob). Open apply_url with browser tools (mcp__claude-in-chrome__navigate + mcp__claude-in-chrome__get_page_text) and read the rendered form. If browser tools are unavailable, extract whatever you can from form_text, then try WebFetch on apply_url.
• Still unreadable — record this explicitly in application-questions.md with the apply_url so the user can open it manually. Never silently skip the form.

Classify each question into three buckets (used in Step 9):

1. Essay/motivation questions ("Why do you want to work here?", "Describe a project…") — these need drafted answers.
2. Factual questions (salary expectation, notice period, visa/work authorization, start date, location, years of experience) — answer from the profile where possible.
3. Standard fields (name, email, phone, CV/resume upload, LinkedIn URL) — list them so nothing is a surprise, but no draft needed.

Step 3 — Read Kristian's Profile

Read the full candidate profile before scoring:

.claude/skills/generate-application/references/kristian-profile.md

Also read the scoring matrix and tone/voice reference:

.claude/skills/generate-application/references/scoring-matrix.md
.claude/skills/generate-application/references/tone-and-voice.md

Step 4 — Score and Analyze

Apply the multi-dimensional scoring matrix (see references/scoring-matrix.md).

Produce for analysis.md:

• Overall score (%)
• Dimension-by-dimension breakdown with reasoning
• Positioning frame (Lead/Principal Engineer, Engineering Manager, Head of, Staff Engineer)
• Requirement → proof point mapping table (every listed requirement matched or flagged as a gap)
• Honest gap assessment
• Recommendation: GO / STRETCH / PASS

If score < 60% (PASS), tell the user and stop — do not generate a cover letter or CV data for a role with poor fit.

Step 5 — Determine Output Slug

Derive folder name:

• company: lowercase, no spaces, no punctuation (e.g. iris, deepmind)
• role-slug: kebab-case from role title (e.g. tech-lead, staff-engineer)
• date: today's date as YYYY-MM-DD
• Full path: applications/[date]-[company]-[role-slug]/

The applications/ directory lives in the project root. Create it (and the slug subfolder) if it doesn't exist.

Step 6 — Save job-posting.md

Write the raw parsed job description to applications/[slug]/job-posting.md. Include the apply_url from Step 2 at the top.

Step 7 — Save analysis.md

Write the full analysis (Step 4 output) to applications/[slug]/analysis.md.

Step 8 — Generate cv-data.js

Generate a targeted CV data file for this role. This is a CommonJS module that exports an object matching the schema in src/_data/cvIris.js (the canonical example of a targeted CV variant — use this as the schema reference, not src/_data/cv.js).

Apply tone and voice guidance from references/tone-and-voice.md when writing the profile statement and employment bullets.

Reframe employment bullets, profile statement, and skill rankings to match:

• The role's tech stack (prioritise skills they listed first)
• The positioning frame chosen in Step 4
• Concrete proof points with metrics from references/kristian-profile.md
• Honest language — do not invent technologies or claim expertise not in the profile

Save to applications/[slug]/cv-data.js.

Also print instructions for the user to integrate it:

1. Copy cv-data.js → src/_data/cv[Company].js (camelCase, e.g. cvDeepMind.js, cvCrossref.js)
2. Rebuild: bun run build

Step 8b — Render cv.pdf

Render the targeted CV to a PDF straight from cv-data.js — no site build required. The script loads the same cv.njk layout/template, applies print styling, forces the Education section to start on page 2, and auto-shrinks the content to fit two A4 pages:

node scripts/cv-to-pdf.js "applications/[slug]/cv-data.js"

Output lands next to the data file as applications/[slug]/[slug].pdf (filename matches the folder). The script prints the fit scale and final page count.

• If it warns still N pages at the … floor, the CV data is too long — trim bullets in cv-data.js and re-run rather than shipping 3+ pages.
• Layout is tunable via flags (--margin-top N mm, --section-gap N px, --pages N, --scale N, --no-fit); run node scripts/cv-to-pdf.js with no args for the full list. Defaults produce the standard two-page layout — only pass flags when a specific CV needs it.

Step 9 — Draft application-questions.md

Write applications/[slug]/application-questions.md covering every question found in Step 2, grouped by bucket:

# Application Form: [Role] at [Company]

**Apply URL**: [apply_url]
**Form source**: api / apply-page / manual (browser) / unreadable

## Essay & Motivation Questions
### Q: [question text] (required, max N words if stated)
[Drafted answer — tone-and-voice rules, concrete proof points with metrics,
respect any stated word/character limit, honest.]

## Factual Questions
| Question | Answer |
|----------|--------|
| [e.g. Notice period] | [from profile, or `[FILL ME: notice period]`] |

## Standard Fields (no draft needed)
- Name, email, CV upload, ...

Rules:

• Draft essay answers with the same voice rules as the cover letter (references/tone-and-voice.md) — no em dashes, no invented experience.
• For yes/no screeners, answer honestly from the profile. If an honest answer is a knockout risk (e.g. "Do you have X?" and the profile says no), flag it prominently rather than fudging it.
• For personal facts not in the profile (salary expectation, notice period, visa status, earliest start date), insert a [FILL ME: …] placeholder — never guess.
• If the form could not be read at all, this file must still exist, containing the apply_url and a note telling the user to check the form manually before submitting.

Step 10 — Generate Cover Letter via Agent

Dispatch the cover-letter-writer agent with:

• The job posting content
• The analysis (positioning frame, top 3 proof points, gaps)
• The company name and role title
• Any essay questions from Step 2 that overlap with cover letter territory (so the letter and the form answers complement rather than repeat each other)

The agent returns a greeting line, a two-paragraph cover letter body in a warm, helpful, audience-first tone, and a concise close line (for example, "Sincerely,"). It avoids em dashes in output and does not force a first-90-days commitment. Save to applications/[slug]/cover-letter.md.

Invoke using the Agent tool with subagent_type matching the cover-letter-writer agent. Pass all context in the prompt.

Step 11 — Write README.md

Write applications/[slug]/README.md using the template in references/output-structure.md. Status should be Draft by default. Include the apply_url and list any [FILL ME] placeholders still open.

Step 12 — Present Summary

Tell the user:

• Score and recommendation
• Folder path where files were saved
• Which files were generated (including [slug].pdf and its final page count)
• The apply deep link, how many form questions were found, and any [FILL ME] placeholders that need their input before submitting
• How to integrate cv-data.js into the site
• One-line suggested email subject line for the application

Reference Files

• references/kristian-profile.md — Full candidate profile with proof points and tech stack
• references/scoring-matrix.md — Scoring dimensions, thresholds, and proof point language
• references/output-structure.md — Folder layout, file formats, and integration steps
• references/tone-and-voice.md — Voice rules for cover letter, CV profile, and employment bullets

Scripts

• scripts/fetch-application-form.sh <job-url> — Discover the apply deep link and extract application form questions ({url, ats, apply_url, source, needs_browser, questions, form_text} JSON). Handles Greenhouse, Lever, Ashby, Workable, SmartRecruiters, Factorial, and generic career pages. needs_browser: true signals the form must be read with browser tools.
• scripts/cv-to-pdf.js <cv-data.js> [out.pdf] [flags] — Render a cv-data.js to a two-page A4 PDF via the site's cv.njk template and puppeteer, no Eleventy build. Auto-fits to --pages (default 2). Lives in the project root scripts/, not the skill dir.

Notes

• Always read the full posting before scoring — do not assume from job title alone
• Be honest about gaps in analysis.md — authenticity over overselling. But keep gap admissions OUT of the cover letter itself; the letter sells fit, it does not disclaim it (see references/tone-and-voice.md)
• The cover letter body must be exactly 2 paragraphs. The agent enforces this.
• The package is not complete without application-questions.md — either with the form's questions answered, or with an explicit note that the form is behind a login/JS wall and must be checked manually
• If the job URL redirects or is behind a wall, ask the user to paste the posting text directly