--- name: resume-ats-linkedin-optimizer description: >- Build, rewrite, tailor, and audit a world-class resume/CV and LinkedIn profile that wins interviews — optimized at the same time for ATS parsing (Workday, Greenhouse, Lever, Taleo, iCIMS, Ashby), fast recruiter skim-reading, and LinkedIn recruiter search. Use when the user wants to create, improve, fix, or review a resume or CV; tailor a resume to a specific job description; run an ATS check or audit; write or strengthen achievement bullet points; quantify accomplishments; optimize a LinkedIn headline, About, skills, or experience; write a tailored cover letter; or generally maximize their chances of getting hired. Governing rule: zero fabrication — never invent facts, employers, dates, titles, or numbers. license: MIT --- # Resume + ATS + LinkedIn Optimizer Turn a person's real experience into the strongest possible resume, CV, and LinkedIn profile — one that survives every major ATS parser, wins the 7-second recruiter skim, and ranks in LinkedIn recruiter search. This skill encodes what university career centers (Harvard, MIT), ex-FAANG recruiters, eye-tracking research, and the actual behavior of production ATS agree on, and resolves the places where popular advice is wrong or outdated. ## The prime directive: zero fabrication **Never invent, inflate, or guess a fact.** No fabricated employers, titles, dates, degrees, or metrics. This is non-negotiable and overrides every other goal — a resume that lies fails the moment a recruiter or reference check catches it, and it puts the user's career at risk. What this means in practice: - Every number, employer, title, and date must come from the user or a document they gave you. If you don't have it, ask — don't fill it in. - Metrics the user *has but didn't write down* are fair game to **excavate** (Phase 2). Estimates are allowed **only** when the user confirms them and they are hedged ("~", "approximately", "roughly"). See `assets/metric-proxies.md`. - You may sharpen *language* freely (weak verb → strong verb, task → outcome, vague → specific). You may not sharpen *facts*. - When you reword something, preserve its truth. "Helped build" does not become "single-handedly architected." If the user pushes you to fabricate, refuse and explain the risk, then offer the honest alternative (excavate a real number, reframe, or omit). ## Mental model (get this right, everything else follows) A modern hiring pipeline is **two stages, and parsing is the gate before everything**: 1. **Parse + store.** The file is flattened to plain text and mapped into fields (name, contact, roles, dates, skills). If the parser can't read something, it is invisible to *every* downstream step, including any AI layer. 2. **Search + rank + (increasingly) AI-summarize.** Recruiters run keyword searches over the parsed database; an LLM layer may summarize and score on top. Humans still make the decision. The load-bearing truth: **ATS almost never auto-reject** (≈92% of recruiters say theirs don't). The "ATS rejects 75% of resumes" stat is a myth from a defunct 2012 vendor. What feels like bot-rejection is human overload (hundreds to thousands of applicants) plus knockout questions. So optimize for **(a) clean parsing and (b) a fast, convincing human skim** — not for gaming a rejection algorithm that mostly doesn't exist. Details + myths: `references/07-myths-and-truth.md`. ## The single source of truth: the Master Profile Do not tailor by rewriting from scratch each time. Maintain a **Master Profile** — an exhaustive, private store of everything the user has ever done (every role, bullet, metric, skill, project). Every resume and LinkedIn variant is then a **selection and reframing** of that store, never an invention on top of it. This is what makes tailoring fast, consistent, and truthful. - If the user has no Master Profile yet, build one first (`assets/master-profile-template.md`). - If they do, load it and update it with anything new before tailoring. - Full workflow and the role "visibility" system: `references/05-master-profile.md`. ## Modes Detect which the user wants (ask if unclear) and follow the phases below. | Mode | Trigger | Phases | |---|---|---| | **Build** | "make/write my resume", from scratch or from raw material | 0 → 6 | | **Improve** | "fix / make my resume better", no specific job | 0,1,2,3,5,6 | | **Tailor** | "tailor my resume to this job" + a job description | 0,1,4,2,3,5,6 | | **Audit** | "check / score my resume", "is this ATS-friendly" | 0,3,4,6 (report only) | | **LinkedIn** | "optimize my LinkedIn" | 0,1,7 | | **Apply/skip** | "should I apply to this?" | 0,4 → `references/03-keywords-and-tailoring.md` gate | | **Cover letter** | "write a cover letter" | 0,4 → `references/08-cover-letters.md` | ## Workflow Work through phases in order for the chosen mode. **Gate each phase**: do not move on until its exit condition is met. Load the linked reference *when you reach that phase* — don't front-load everything. ### Phase 0 — Intake & Master Profile - Establish the goal: target role/industry, seniority, and whether there's a specific job description. Determine career stage (new grad / mid / senior / career-changer) — it drives section order and emphasis. - Gather raw material: existing resume, LinkedIn export, job history, or a conversation. Load into / create the **Master Profile**. - **Exit when:** you have a Master Profile (even a rough one) and a clear target. ### Phase 1 — Metric Excavation & fact-gathering - The highest-leverage step. Most people undersell themselves. For each significant role/bullet, ask targeted questions to surface real numbers: scope (team size, budget, users, volume), change (before → after, %), time/speed, money, and quality. Use `assets/metric-proxies.md` to *prompt* plausible ranges — but every number must be **confirmed by the user**. - Also collect: domain/context (the "secret weapon" — "web apps" → "fintech card-issuance platform"), tools/tech, and outcomes. - **Exit when:** each headline achievement has real, user-confirmed substance. ### Phase 2 — Content engineering (bullets, verbs, summary) - Rewrite every bullet to **Action verb + what you did + quantified impact + (how/context)** — the XYZ / Google formula ("Accomplished X, measured by Y, by doing Z"). Lead with impact, not task. - Use strong, *plain* verbs (led, built, designed, cut, shipped, grew); kill "responsible for / helped / worked on" and buzzword-inflated verbs. Full guidance, verb bank, and before/after examples: `references/02-content-and-bullets.md` and `assets/action-verbs.md`. - Run the **de-slop pass**: remove AI-tells and filler (`assets/ai-slop-blacklist.md`). - Write the professional summary (if ≥1–2 yrs experience) or objective (new grad / career-changer). - **Exit when:** bullets are outcome-led, quantified where honest, and slop-free. ### Phase 3 — Structure & ATS-safe formatting - Enforce the non-negotiable ATS format: **single column; no tables, text boxes, columns, headers/footers, graphics, or icons; standard section headings; standard bullets; web-safe font 10–12pt; contact info in the body.** - Order sections by career stage; reverse-chronological; dates as `Month Year – Month Year` + "Present". Full rules: `references/01-ats-and-formatting.md`. - **Exit when:** structure passes the ATS-safe checklist in that reference. ### Phase 4 — Keywords & tailoring (Tailor/Audit/Apply modes) - Analyze the job description: extract and tier keywords (required vs preferred), find the most-emphasized terms, run a gap analysis vs the Master Profile. - Mirror the JD's **exact** hard-skill wording (ATS often don't resolve synonyms); include acronym + spelled-out pairs ("SEO (Search Engine Optimization)"). Place each key term in ≥2 places (skills + a real bullet). **No stuffing, no hidden text** — both backfire. Use `scripts/resume_tools.py coverage` for a labeled keyword-coverage estimate. Details + apply/skip gate: `references/03-keywords-and-tailoring.md`. - **Exit when:** required keywords the user genuinely qualifies for are present and natural. ### Phase 5 — Render - Prefer the **LLM-writes-JSON, code-renders-doc** pattern: put content in `assets/resume.schema.json` shape, then `python scripts/resume_tools.py render --resume resume.json --out resume.pdf` (or `.docx`) produces a clean, single-column, ATS-safe file with selectable text. This keeps layout deterministic and parseable. Text-based PDF is the safe default for submission; keep `.docx` for portals that request Word. - File name: `Firstname-Lastname-Resume.pdf`. - **Exit when:** a rendered file exists and its text is selectable/copyable. ### Phase 6 — Evaluate - Run the scoring rubric and the **truthfulness invariants** (no section dropped, no employer/date/title that wasn't in the source, identity untouched): `python scripts/resume_tools.py validate` and see `references/06-scoring-and-evaluation.md`. - Do the **skim test**: in 7 seconds can you find name, current title+company, and the top achievement? Is the strongest material in the top third? - Report a **keyword-coverage estimate** (explicitly labeled an estimate, never a fake "100/100 ATS score"). - **Exit when:** invariants pass and the skim test succeeds. ### Phase 7 — LinkedIn (LinkedIn mode) - Port the same keyword spine into LinkedIn, adapting for its algorithm and human, in-feed reading: headline (220 chars, front-loaded, keyword-rich), About (hook in the first ~300 chars, first-person story, CTA), experience (conversational + media), pinned top-3 skills, All-Star completeness. Per-field limits, formulas, and checklist: `references/04-linkedin.md`. - **Exit when:** every field has a concrete recommendation and the profile hits All-Star. ## Reference & asset map Load references on demand as you reach the phase that needs them. - `references/01-ats-and-formatting.md` — how each ATS parses, what breaks it, format rules, headings, dates, PDF vs DOCX, file naming, length. - `references/02-content-and-bullets.md` — XYZ/CAR formulas, quantification, summary vs objective, recruiter skim behavior, red-flag handling, tense/voice. - `references/03-keywords-and-tailoring.md` — JD analysis, keyword tiering, coverage scoring, apply/skip gate. - `references/04-linkedin.md` — algorithm, per-field limits + formulas, skills, featured/recommendations/URL/banner/photo, resume↔LinkedIn diffs. - `references/05-master-profile.md` — the source-of-truth doc + visibility system. - `references/06-scoring-and-evaluation.md` — rubric, truthfulness invariants, skim test. - `references/07-myths-and-truth.md` — myths to stop repeating; what's actually true. - `references/08-cover-letters.md` — when to write one, structure, rules, template. - `assets/master-profile-template.md`, `assets/resume.schema.json`, `assets/example-resume.json`, `assets/action-verbs.md`, `assets/metric-proxies.md`, `assets/ai-slop-blacklist.md`. - `scripts/resume_tools.py` — `render` (JSON→ATS-safe .docx), `validate` (truthfulness invariants), `coverage` (JD keyword-coverage estimate), `deslop` (flag AI-tells/filler). `python scripts/resume_tools.py --help`. ## Hard rules (never violate) 1. **Zero fabrication.** See prime directive. 2. **Single column, no tables/graphics/headers-footers.** Parsing gate. 3. **Standard section headings and reverse-chronological order.** 4. **Exact-match JD keywords, placed naturally — never stuffed or hidden.** 5. **Quantify honestly.** Estimates only when user-confirmed and hedged. 6. **No fake "ATS score."** Report a labeled keyword-coverage *estimate* only. 7. **Plain strong verbs over buzzwords.** Run the de-slop pass on every draft.