/resume-studio

name: resume-studio
description: Build, tailor, and maintain job-winning resumes with a persistent career memory. Use this skill WHENEVER the user mentions a resume or CV in any capacity — creating one, updating one, tailoring one for a job description, reviewing/critiquing one, converting a LinkedIn profile into a resume, preparing a job application, or sharing details about their work history that should be remembered. Also trigger when the user shares a JD and wants application materials, says "help me apply for this job", wants to fix or rewrite resume bullets, or corrects how their resume should be written (style/framing preferences). Even a small resume edit should go through this skill, because it maintains the user's career corpus and style rules across sessions.

Resume Studio

Build the user a resume that wins interviews, from a career memory that compounds over time.

Why this skill works the way it does

Three research-backed convictions drive everything below (full evidence: `references/resume-playbook.md`):

1. **Every resume has two readers.** An ATS parser that converts the document into structured database fields, and a human who skims it in seconds against a stack of hundreds. The parser needs clean, single-column, standard-labeled text. The human needs specific, quantified, verifiable accomplishments. Optimize for both; "beat the ATS" keyword tricks are folklore — ATS platforms almost never auto-reject, humans and knockout questions do. 2. **The bottleneck is elicitation, not writing.** Users under-report their own impact. The job is to extract what they actually did — metrics, scope, stories — not to polish what they happened to remember. People are bad at answering open-ended questions but great at correcting a draft, so always draft first and interrogate the gaps. 3. **Fabrication is the cardinal sin.** In the AI-application era, hiring teams actively verify claims in interviews. NEVER invent a number, title, date, or achievement. Every quantified claim in any output must trace to something the user said or uploaded. If evidence is missing, ask — or write the bullet without the number. When summarizing or rephrasing, stay within what the evidence supports.

Workspace: where state lives

This skill is stateless; the user's data lives in a `resume-studio/` folder (create it in the user's connected folder or working directory on first use — ask where they want it if a device folder is connected):

resume-studio/
├── career-corpus.md        # single source of truth for their professional history
├── style-rules.md          # learned writing/formatting preferences (a lint file)
├── changelog.md            # dated log of what was learned/changed and why
└── applications/
    └── YYYY-MM-company-role/   # one folder per application
        ├── jd.md               # the job description + any HM notes/posts
        ├── research.md         # company/role research notes
        ├── resume.md           # the tailored resume (markdown master)
        ├── resume.docx / .pdf  # rendered outputs
        └── delta-report.md     # what was tailored and why + honest gap report

File formats are specified in `references/corpus-schema.md` — read it before creating or editing `career-corpus.md` or `style-rules.md`.

**On every invocation, before doing anything else:** 1. Locate the `resume-studio/` folder (check the connected device folder and working directory; ask if not found and ambiguous). 2. Read `career-corpus.md` and `style-rules.md` if they exist. Style rules are binding on all output you produce — treat them like a linter. 3. Skim the 2–3 most recent `applications/` folders. If the user corrected something there, do not repeat the mistake — repeated corrections are the single fastest way to lose a user's trust.

Then pick a mode:

| Situation | Mode | |---|---| | No corpus exists, or user shares new career material | **BUILD** | | Corpus exists + user has a target job/JD | **TAILOR** | | User corrects style/facts, or asks to "sync"/"update my profile" | **SYNC** (also runs implicitly, always) | | User asks for a review of an existing resume | **TAILOR** (critique path: score their resume against the playbook + JD before rewriting) |

Modes chain naturally: a first-time user with a JD gets a fast BUILD, then TAILOR in one session.

Mode: BUILD — the career corpus

Goal: a corpus so rich that any future resume is an editing job, not a writing job. The corpus is NOT a resume — it holds far more than will ever fit on one page, including material that only matters for specific job types.

**Step 1 — Ingest everything offered.** Resumes (all versions — old ones contain forgotten detail), LinkedIn profile/PDF, bio pages, portfolio links, and free-form dumps. Proactively ask for the high-signal sources users never think to share, in this order of value: performance reviews and self-assessments, promotion/appraisal docs, brag docs, launch emails, OKR docs, LinkedIn recommendations, the JDs of jobs they *held*, offer letters. Invite voice-note-style rambling: "talk for five minutes about the project you're proudest of" out-performs typed Q&A. Don't block on these — work with what arrives.

**Step 2 — Draft the corpus immediately, annotated.** Convert everything into `career-corpus.md` per the schema (stable anchors per role/project, provenance per claim). Mark weaknesses inline where they occur:

• `[NO METRIC]` — accomplishment with no quantification
• `[VAGUE]` — can't tell what the user specifically did vs their team
• `[THIN]` — role/period with almost no content

Show the user the draft early. The annotations do the motivational work — visible holes make people talk.

**Step 3 — Interview against the gaps.** Read `references/question-bank.md` for techniques. Rules of engagement:

• Ask 3–5 questions per round, highest-value gaps first (recent roles and target-relevant skills before ancient history). Use the AskUserQuestion tool where it exists, with free-tex