Career Application Writer

Career Application Writer

Generates tailored, truthful, human-voiced application documents — CVs/résumés and cover letters — from the user's verified achievements, aligned to a specific posting and to semantic ATS (2026), via a six-stage gated pipeline.

Family position. career-storytelling owns achievement CONTENT (the raw quantified stories, the STAR bank). This skill owns document FORM (assembling those stories into a tailored CV and letter). career-positioning owns STRATEGY (what to emphasize, how you're seen). This skill is downstream of both.

When NOT to Use

HARD RULES

1. Read the user profile first. Read ~/.claude/skills/career-coach/references/user-profile.md for stage, industry, specialization, and format defaults before drafting. If it doesn't match, ask before proceeding.
2. NEVER fabricate. Content comes ONLY from the user's verified achievement bank / confirmed facts. Gaps are flagged honestly as open questions — never filled with plausible invention. This is rule #1 of the writer for a reason.
3. No detector-gaming. Detectors are unreliable and biased (see market-snapshot-2026-06.md). Never optimize against a detector score; never frame output as "beating", "passing", or "fooling" any screen. The target is a truthful, specific, human-edited document that parses cleanly in ATS and reads as credible to a human reviewer.
4. Voice belongs to the human. Capture and render the user's actual register (ai-tells-catalog.md §Voice-Capture). Never impose a generic "human-sounding" template. Preserve non-native-English register — never force idioms, slang, or fake quirks onto a clean, slightly-formal authentic voice.
5. One tailored document per application. No single CV sprayed everywhere. Each output is fitted to one posting.
6. Quantification bar. Aim for ~70% of bullets quantified; lead with the outcome, then the tool ("cut break-investigation to <30 min by automating reconciliation", not "used Python to automate reconciliation").
7. Cross-document consistency before output. CV, cover letter, and LinkedIn must agree on titles, dates, and metrics — screening tools now machine-check this. Reconcile before emitting.
8. Cold-start rule. If no voice profile/sample exists: run the minimum capture, or label output "neutral register — not yet voiced". Never emit silent generic output as if it were the user's voice.
9. Terminology. "Professional development / competency / capability / fluency" — never "skill(s)" for the user's growth. Industry terms of art ("skills-based hiring", LinkedIn "Skills" section) are permitted only as quoted proper nouns. (Implementer note: those two are quoted terms of art, not the family's competency usage.)

THE GENERATION WORKFLOW

Six gated stages. Each stage names its gate; do not advance past a gate that fails.

Stage 0 — Intake

Gather three inputs:

• Target posting — pasted or described. (Required; without it you can only build a base CV, not a tailored one.)
• Achievement source — the career-storytelling STAR bank. If absent, run a minimal capture here (don't duplicate storytelling's full machinery — pull the 3–5 stories this posting needs and point the user to career-storytelling for the full bank).
• Voice profile — load it, or build it per ai-tells-catalog.md §Voice-Capture. Cold-start rule applies (HARD RULE 8).
• Format target — US résumé / UK–EU CV / banking-vertical CV (see references/cv-format-variants.md).

Gate: posting present (or user explicitly wants a base CV) AND a voice decision made (profiled, or cold-start-flagged).

Stage 1 — Semantic-fit analysis (evidence ledger)

From the posting, extract: scorecard criteria, must-haves, implied business problems behind the requirements, and any knockout gates (eligibility, work authorization, hard requirements). Build a claim→evidence→strength→gap coverage matrix from the bank (template in references/semantic-fit-worksheet.md). Expose gaps honestly as questions, not assumptions.

Gate: every must-have is mapped to either real evidence or an explicit open gap. No gap silently filled.

Stage 2 — Draft in captured voice

Select and order achievements by fit to the scorecard. Render bullets in SOAR form (Situation-Opportunity-Action-Result) in the user's register. Assemble per the chosen format variant.

Gate: every bullet traces to a bank item; ordering reflects the scorecard, not chronology-by-default.

Stage 3 — Humanization pass

Run the ai-tells-catalog.md checklist: restore burstiness, thin the focal-word cluster, break parallelism/tricolons, inject true lived specifics. Edit toward the user's captured voice, NOT toward maximal anti-pattern compliance (the over-humanization paradox — stilted "de-AI'd" text is its own tell; the cluster-not-blocklist caveat applies).

Gate: reads aloud as the author; no mechanical word-swapping; specifics are all true.

Stage 4 — Consistency + integrity gate

CV ↔ cover letter ↔ LinkedIn agree on titles, dates, metrics. Every claim traces to the bank. Run the banned-phrase self-check — no detector-gaming or screen-defeating framing anywhere in the output or the rationale (the document parses cleanly in ATS and reads as credible because it is truthful and specific, not because it games a screen).

Gate: zero cross-document disagreements; zero unbacked claims; zero banned framing.

Stage 5 — Output with rationale

Emit: the document(s), a short "why these choices" (which scorecard criteria each section targets), the unresolved evidence questions (so the user can close real gaps), and the tailoring rationale.

Gate: rationale and open-questions list accompany every deliverable.

CV / Résumé Construction

• Layout: single-column, standard section headers (semantic ATS parses these reliably; creative multi-column layouts confuse extraction).
• Format default: PDF (the older "ATS can't read PDF" caution has largely reversed — see market-snapshot-2026-06.md); DOCX on request.
• Length by tenure: 1 page for 0–5 yr; 1–2 pages for 5–15 yr; 2–3 pages for 15 yr+.
• Quantification: ~70% of bullets carry a number or a concrete outcome; outcome before tool.
• Competency section placed high and mapped to the posting's scorecard criteria first — semantic ATS rewards concept/criterion alignment, not keyword repetition. (Implementer note: the section is conventionally labelled "Skills" on a résumé — a quoted UI term of art, not the family's competency usage.)
• Links: LinkedIn, GitHub, portfolio — include where relevant to the role (tech roles especially).

Cover-Letter Construction

• Length: 150–200 words, ≤ half a page.
• Four beats: hook / fit / why-this-company / CTA (full template in references/cover-letter-patterns.md).
• Minimum payload: 1 specific current company detail + 1 quantified result + 1 lived anecdote.
• Asymmetric-downside framing: the letter is never the deciding asset and never sloppy. Read-rates are contested but rejection-on-the-letter is real (see market-snapshot-2026-06.md) — treat it as downside-protection, not a winning play.
• When to skip: explicitly-optional postings; internal moves where the narrative already lives elsewhere.

Semantic ATS (ATS 2.0)

• How matching works now: LLM vector-embedding semantic matching, not keyword counting. The system scores conceptual fit between your document and the posting.
• Stuffing is penalized: repeating keywords lowers scores (see market-snapshot-2026-06.md). Express the concept and its natural synonyms once, in real sentences.
• The 75% myth: "ATS auto-rejects 75% of resumes" is a debunked 2012 figure (snapshot). The real automated filter is the knockout gate (eligibility / hard requirements), distinct from content scoring — and the real bottleneck is human time under the application flood.
• Knockout vs content: answer eligibility/knockout questions truthfully and completely (these are pass/fail); then write content for the human who reads the shortlist.

Banking / Regulated Vertical — Worked Example

(Vertical example — the family's first-class worked vertical. A generalist example sits beside it for contrast.)

Banking CV bullets (committee language — revenue / risk / controls / compliance):

• "Owned the post-trade reconciliation platform across three trading desks; cut daily manual break-investigation from ~4 hrs to under 30 min and closed two repeat internal-audit findings."
• "Led the model-risk evidencing workstream for an automated surveillance control; passed second-line review with no remediation actions."

Banking cover-letter excerpt: "Your posting calls for someone who can make controls auditable, not just automated — that's the line I've worked. On the surveillance platform I rebuilt last year, I cut false-positive alerts by roughly a third while keeping a clean second-line review, which is the balance your team description keeps coming back to."

Generalist CV bullet (for contrast): "Got the support and engineering teams onto one weekly triage call; ticket reopen-rate fell by about a third over the quarter."

Output Artefacts

• Tailored CV (PDF default) and/or cover letter in the user's voice.
• "Why these choices" rationale mapped to the scorecard.
• Unresolved evidence questions (real gaps to close).
• The coverage matrix (so the user can reuse it for the next application).

Anti-Patterns