career-init

Files (12)

• SKILL.md
• CLAUDE.md · templates/
• PROFILE.md · templates/
• Preferences.md · templates/
• STATE.md · templates/
• Person.md · templates/notes/
• Praise.md · templates/notes/
• Role.md · templates/notes/
• Skill.md · templates/notes/
• Voice Profile.md · templates/notes/
• Win.md · templates/notes/
• resume-master.md · templates/notes/

SKILL.md (first lines)

# /career-init - stand up the Career Brain

Run once. Everything after this is `/career-interview`, `/career-ingest`, `/career-synthesize`, `/career-jobs`, and `/career-apply`.

If `~/Career/STATE.md` (or the configured vault path) already exists, do NOT re-scaffold. Say so, summarise what is already there and what is missing, and hand off to the right command instead.

## The governing idea

The user should never be asked to recite something a machine could have found. Their time is only worth spending on what nothing else can know: what they are proud of, what they hated, what they would not do again, why a decision was hard, what they learned. So the order is always **ingest first, interview second**. Arriving at the interview already knowing their last four jobs is the difference between this feeling like an assistant and feeling like a form.

## Phase 0 - Frame it, briefly

Two or three sentences, no bullet lists, no enthusiasm inflation. Roughly: this builds a local record of what they have done and what people have said about it, so they can see their own value clearly and describe it to other people. It takes about an hour to get somewhere useful, it saves progress as it goes, and they can stop at any point.

Then ask one thing before doing anything else: **what is prompting this?** Between a jobs, a review cycle, a promotion case, a rate negotiation, quiet dissatisfaction, and "I just do not know how to describe what I do", the answers differ enormously in what matters. Record it in `STATE.md` as `motivation`. Do not skip this to seem efficient.

## Phase 1 - Environment check

Check what is actually available and report it plainly as a short table: source, available yes/no, what it gives you.

1. **Granola MCP** - try `list_meetings` with a small limit. This is the single highest-yield source for unsolicited praise and for evidence of what they actually spend time on. If missing, point them at the Granola MCP setup and note that everything still works without it.
2. **Claude in Chrome** - call `tabs_context_mcp`. Needed for LinkedIn, which is the fastest route to a full role history. If missing, they can paste their LinkedIn profile text instead and lose very little.
3. **Gmail and Google Calendar connectors** - a trivial search to confirm. Gmail carries praise threads and review cycles, Calendar reveals the real collaboration graph.
4. **GitHub** - `gh auth status`. Only matters if they write code.
5. **Slack, Drive** - note availability, do not chase.

Say honestly which gaps cost the most. Do not pretend a missing source is fine if it is not, and do not stall the setup over one. Offer to continue and add sources later.

## Phase 2 - Vault location and scaffold

Default `~/Career`. Offer it, accept an alternative, and confirm before writing.

Create this structure, copying the files from the `templates/` directory that sits alongside this skill file. Copy `templates/notes/` into the vault as `Templates/`, so that every later command reads templates from the vault rather than from wherever this skill happens to be installed. Installation paths differ between the plugin, the standalone skill, and the shell installer; the vault does not.

```
<vault>/
  CLAUDE.md            <- conventions, loaded into every session in this vault
  STATE.md             <- progress, gaps, open questions, intensity dial
  PROFILE.md           <- the living positioning one-pager (stub until synthesize runs)
  Preferences.md       <- likes, dislikes, energy, dealbreakers
  Templates/           <- note templates, copied in at setup
  Incoming/            <- drop zone for resumes, cover letters, reviews, anything
    processed/
  Roles/               <- one file per job, contract, or engagement
  Wins/                <- one file per achievement
  Praise/              <- one file per piece of external evidence
  Skills/              <- one file per capability, linked to evidence
  People/              <- collaborators, advocates, references
  Voice/
    Voice Profile.md
    Samples/
  Sources/             <- raw ingested material (gitignored)
  Jobs/                <- leads and applications (later phase)
  Assets/
    resume-master.md   <- single source of truth for resume bullets
```

Write a vault `.gitignore` containing `Sources/`, `Incoming/`, and `.DS_Store`. Then `git init` and make one initial commit. Explain in one line why raw sources are excluded: transcripts and private messages should not end up in a repo she might later push.

Do NOT create a remote. Do NOT push.

## Phase 3 - Set the intensity dial

Explain the dial in two sentences and let them pick a starting point. Write it to `STATE.md` as `question_intensity`.

- `high` - ask freely, chase interesting threads, longer sessions
- `normal` - default, batches of three to five questions, stop when a topic is covered
- `low` - only ask what is blocking a specific piece of work
- `paused` - no unprompted questions at all

Tell them plainly that "ease off" or "that's enough" changes this setting permanently rather than for one turn, and "ask me more" turns it back up. This matters: people get question fatigue at very different points and the common failure is that a tool keeps interrogating someone who has mentally moved on to applying for jobs.

## Phase 4 - Seed the record (ingest before interviewing)

Do the cheap, high-yield passes now, in this order. Each is a delegated call to the ingest logic; run them, summarise briefly, and move on. Do not narrate every file.

1. **Incoming folder.** Before anything else, tell them what to put in `Incoming/`: old resumes, cover letters they have sent, performance reviews, 360 feedback, talk abstracts, conference feedback, screenshots of praise, messy notes. Give them a moment to drag files in. Then run `/career-ingest incoming`. This is the highest-value five minutes in the whole setup: a single old resume seeds most of the role history, and a single old cover letter seeds the voice profile.
2. **LinkedIn** via `/career-ingest linkedin`, if Chrome is available.
3. **Granola** via `/career-ingest granola`, if available. Cap the first pass at the last six months so setup does not stall.

If a pass finds nothing, say so and continue. After all three, report what you now know in a compact form: roles found, date coverage, praise items, obvious holes. This report is what makes the interview feel earned.

## Phase 5 - First interview

Hand off to the `career-interview` logic. On this first run, three things are specific to init:

- **Open by reflecting back, not by asking.** Summarise what you have gathered about them and invite correction first. People correct more readily than they volunteer, and every correction is high-quality signal.
- **Cover the spine before the detail.** Get every role in `Roles/` at least stubbed with dates, employer, title, and scope before going deep on any one of them. An incomplete spine distorts everything downstream.
- **Then go deep on the two or three roles that carry the most weight** - most recent, longest, or the ones they light up about.

Respect the intensity dial from the very first batch.

## Phase 6 - First synthesis

Run the `career-synthesize` logic. Produce `PROFILE.md` even if the evidence is thin, and mark it clearly as a first draft with a confidence note per section. Seeing a rough profile with visible holes motivates far better than a request for more input does.

## Phase 7 - Handoff

Close with, in prose not a wall of bullets:

1. What now exists, in numbers: roles, wins, praise items, skills.
2. The three biggest gaps, and which command closes each.
3. What to do next, as one specific suggestion rather than a menu.
4. That everything is saved and resumable, and how to change the question volume.

Then stop. Do not roll straight into another interview round.

From bendechrai/career-brain, MIT.

More from bendechrai/career-brain