Onboard

When to use

On a fresh clone of the repo, or any time the user explicitly invokes /onboard. Also when the user says things like "I just cloned this", "set me up", "first time", "how do I start".

If the user is mid-task on something else (writing a resume, analyzing a Job Description), do not silently redirect them into onboarding. Confirm first.

What it produces

• $PERSONAL/ directory bootstrapped with career/ and applications/ subfolders.
• Four core personal docs populated (or skipped at the user's choice) in $PERSONAL/career/:
• personal-info.md
• goals.md
• impact-doc.md
• brag-doc.md
• A short readout of what was filled, what was deferred, and what tools (if any) are missing.

Nothing else. No git changes. No app launching. No resume generation.

Workflow

Run these phases in order. Each phase is resumable: re-running /onboard after partial setup detects what already exists and picks up where the user left off. Confirm before overwriting any file that already has real content (not template stubs).

Phase 0: Resolve $PERSONAL and detect state

1. Resolve $PERSONAL and $LEETCODE_BANK per CLAUDE.md (both are siblings of the main repo).
2. Check what already exists:

• Does $PERSONAL/ exist?
• Does $PERSONAL/career/ exist?
• Does each of personal-info.md, goals.md, impact-doc.md, brag-doc.md exist?
• For each existing file: is it still a template stub (contains the original <!-- comments --> and empty fields) or does it have real content?
• Does $LEETCODE_BANK/ exist (for Phase 2.5)?

1. Print a one-line state summary, e.g.:

   Setup status: $PERSONAL exists, personal-info filled, goals stub, impact-doc and brag-doc missing. $LEETCODE_BANK missing.
   Resuming from goals.

1. If everything is already filled with real content, ask the user whether they want to re-interview any section or just exit.

Phase 1: Prerequisite checks

Run these silently and only report the missing ones. Don't list things that work.

For each missing item, print a one-line fix:

• pandoc → brew install pandoc / winget install pandoc / apt install pandoc.
• Python deps → pip install python-docx reportlab requests.
• Git config → git config --global user.name "Your Name" etc.

If everything is present, say so in one line ("Prereqs: all good") and move on. Do not block onboarding on missing tools. They're only needed when the relevant skill is later invoked.

Phase 2: Bootstrap $PERSONAL

If $PERSONAL/ or its subfolders don't exist, create them and copy the templates that aren't already in place:

$PERSONAL/career/
$PERSONAL/applications/
$PERSONAL/career/personal-info.md   (copy from career/personal-info.template.md if missing)
$PERSONAL/career/goals.md           (copy from career/goals.template.md if missing)
$PERSONAL/career/impact-doc.md      (copy from career/impact-doc.template.md if missing)
$PERSONAL/career/brag-doc.md        (copy from career/brag-doc.template.md if missing)

Never overwrite an existing file at this phase. Templates only land in slots that are empty.

Phase 2.5: Bootstrap $LEETCODE_BANK

The coding-prep skill's company-lookup flow reads from a read-only sibling clone of an upstream open-source repo. Resolve $LEETCODE_BANK per CLAUDE.md (sibling of the main repo, named leetcode-companywise-interview-questions/).

1. If `$LEETCODE_BANK` does not exist: clone it shallow.

   git clone --depth 1 https://github.com/snehasishroy/leetcode-companywise-interview-questions.git "$LEETCODE_BANK"

Tell the user what you're doing in one line ("Cloning company-tagged question bank to $LEETCODE_BANK") before running. The clone is several MB and takes a few seconds. Confirm success in one line.

1. If `$LEETCODE_BANK` already exists: fast-forward to upstream.

   git -C "$LEETCODE_BANK" pull --ff-only

Silent if up-to-date. If it errors (uncommitted changes, diverged), tell the user and skip; do not force-update someone else's clone.

1. If `git` is missing or the clone fails (no network, etc.): report in one line and continue. The bank is optional infrastructure; everything else still works. Suggest the user re-run /onboard later or clone manually.

Never modify files inside $LEETCODE_BANK. It is upstream-owned. The only writes here are git pull --ff-only to refresh.

Phase 3: Set expectations

Before starting the interview, tell the user (once, briefly):

I'll walk you through four documents: personal info, goals, impact doc, brag doc. You can fill any of them in now, defer to later, or paste raw text I'll structure for you. Filling everything now gives the most accurate resumes, fit analyses, and interview prep. Deferred sections will block the skills that depend on them (e.g., resume-builder needs impact-doc).

Then ask: "Want to go through all four, or pick which ones to do now?"

Phase 4: Interview, one doc at a time

For each doc in order, ask the user up front:

personal-info.md — quick form: name, email, phone, location, LinkedIn, GitHub. ~2 min. Fill now, defer, or paste a block I'll parse?

Three branches:

• Fill now (guided): Ask each field one at a time. Keep it tight. If the user already gave a value earlier in the session (e.g., from CLAUDE.md context like email), prefill and confirm rather than re-asking. Write the file when the section is complete.
• Paste: User pastes a block of raw text (e.g., LinkedIn export, an old resume, freeform). Parse it, populate the file, show the result, ask for corrections.
• Defer: Leave the template stub in place. Note this in the final readout. Move on.

Doc-specific guidance:

personal-info.md — simple key/value. Strip the placeholder formatting; write actual values. If the user volunteers links (portfolio, blog), add them as new lines.

goals.md — the template has six sections (current state, target roles, target companies, constraints, what I want, what I'm avoiding). Don't ask all 30+ fields. Ask in groups:

1. "Current role and level?" (title, level, YOE, preferred coding language)
2. "What roles are you targeting?" (titles, level direction, IC vs. EM, domains)
3. "Target companies, tier 1 / tier 2 / no-go list?"
4. "Constraints: location, comp expectation, timeline, dealbreakers?"
5. "What's drawing you to a new role, and what are you avoiding from past roles?"

Skip any group the user says they don't have an answer for yet; leave that section as a comment they can fill in.

impact-doc.md — the most important doc and the longest. Ask:

1. "List your companies in reverse-chronological order, with dates."
2. For each company: "Title progression?" then "What are the 2 to 4 projects you'd want a hiring manager to know about?"
3. For each project, prompt: what it was, your specific role, tech stack, metrics/scope, outcome, what was hardest. One project at a time. Don't batch.
4. After projects, ask about cross-cutting work (mentorship, hiring, on-call leadership, design reviews).

If the user has an existing resume, LinkedIn dump, or old impact doc, offer to ingest it first: "Paste what you have, I'll structure it, then we'll fill the gaps." This is usually faster than cold interviewing.

If the user wants to defer impact-doc, warn them explicitly: "`resume-builder`, `cover-letter-builder`, `star-stories`, and `job-analyzer` all depend on this. They'll work but the output will be generic until you fill it in."

brag-doc.md — four sections (quantified wins, recognition, things shipped, talks/writing/mentorship, failures/lessons). Ask:

1. "Top 5 quantified wins from the last 2 to 3 years? Metric + 1-line context each."
2. "Recognition: promotions, awards, named callouts?"
3. "Failures or hard lessons worth keeping? Useful later for behavioral stories."

Skip "things shipped" and "talks/writing" unless the user volunteers; impact-doc already covers most of that.

Phase 5: Final readout

Print a compact summary:

Setup complete.

Filled:
- personal-info.md
- goals.md (deferred: target companies, dealbreakers)

Deferred:
- impact-doc.md  → blocks resume-builder, cover-letter-builder, star-stories, job-analyzer
- brag-doc.md    → used by resume-builder and interview-prep for quantified bullets

Missing tools: pandoc (needed for .docx output)

Suggested next step:
Fill impact-doc.md when you have ~30 min. Then run /help to see what each skill needs.

Then stop. Don't auto-launch another skill.

Behaviors and edge cases

• CLAUDE.md says "no em dashes." Apply this from the moment of onboarding onward (in the files you write and in your responses). The user has confirmed this preference.
• CLAUDE.md says "work on master." If the user is on a branch when they invoke /onboard, gently flag it. Don't force a switch.
• Templates may have changed since clone. If the template file in the repo is newer than the personal copy and the personal copy is still a stub, prefer the latest template content. If the personal copy has real content, leave it alone.
• Don't ingest external accounts. No reaching out to LinkedIn, GitHub, etc. The user pastes what they want to share.
• Persist incrementally. Write each file as soon as its section is done. Don't hold everything in memory and write at the end; a crashed session shouldn't lose 40 minutes of interviewing.

Anti-patterns

• Don't auto-fill from CLAUDE.md context without confirming. The user's email is visible in the system context but they still get to confirm it before it lands in personal-info.md.
• Don't editorialize about gaps. If a user defers a section or has thin content, note the dependency impact and move on. No "you really should fill this in" lectures.
• Don't validate format pedantically. If the user gives "tc: 320k" instead of a structured field, accept it and write it as-is. The docs are for the user, not for a schema.
• Don't run other skills inside onboard. Onboarding ends at the readout. The user runs /help or a specific skill next.
• Don't ask "are you ready?" between every step. Move through the interview at a brisk pace. Pause only at doc boundaries.
• Don't create $PERSONAL silently if it already exists. Detect, report, and resume.