career-ingest

Files (1)

• SKILL.md

SKILL.md (first lines)

# /career-ingest - turn scattered material into evidence

Argument: `incoming`, `linkedin`, `granola`, `gmail`, `calendar`, `github`, `drive`, `slack`, or `all`. With no argument, read the ingestion log in `STATE.md`, recommend the source with the best yield for the least effort, and ask which to run.

## Rules for every source

1. **Do not send source content anywhere.** Transcripts, DMs, and emails are read locally and stay local. Extracted quotes go in `Praise/`; raw material goes in `Sources/`, which is gitignored.
2. **Attribute everything.** Name, their relationship to the user, date, and source type. Unattributed praise is not evidence.
3. **Never fabricate to fill a template.** A field with nothing behind it stays empty. Empty is recoverable; invented is not.
4. **Mark uncertainty inline** as `[unconfirmed]` rather than dropping it or asserting it. Then add it to the open questions queue so the next interview resolves it.
5. **Log every run** in the `STATE.md` ingestion log with date and coverage window, so re-runs are incremental rather than repetitive.
6. **Report in aggregate.** "Read 84 meetings, found 11 pieces of praise and 3 new collaborators" is useful. A file-by-file narration is not.

## incoming

The highest-yield source, and the first one to run.

If `Incoming/` is empty, tell them specifically what to put there rather than asking whether they have anything: old resumes and CVs, cover letters they have sent (especially ones that got interviews), performance reviews, 360 feedback, talk abstracts, conference feedback forms, workshop retros, screenshots of praise, and any messy personal notes. Then wait.

For each file, extract in two passes:

**Facts.** Roles, dates, titles, projects, numbers, talks, publications. Old resumes are the fastest possible route to a complete spine. Performance reviews are the densest source of attributed praise in existence and are usually sitting in a folder nobody has opened since the day they were written. Where a resume bullet has a number, capture the number and note that it is self-reported unless something corroborates it.

**Voice.** Cover letters, posts, and long-form writing go to `Voice/Samples/` and update `Voice/Voice Profile.md`. Observe sentence shape, vocabulary, how they open, how they claim credit, how they handle gaps, and whether they use humour. Every observation in the voice profile must name the sample it came from. Where the user knows an application's outcome, record it, because a cover letter that got an interview is a better model than one that did not.

Where a resume and their memory disagree on a date or title, trust the contemporaneous document and flag the discrepancy for confirmation.

Move processed files to `Incoming/processed/`. Never delete anything.

## linkedin

Requires Claude in Chrome. Load the browser tools in one `ToolSearch` call, then open a new tab in their existing session so they are already authenticated. Read pages with `read_page` or `get_page_text` rather than screenshots. Go at a human pace. Do not attempt to enumerate large lists exhaustively.

In order of value:

1. **Recommendations received** (`/in/<handle>/details/recommendations/`). Pure attributed praise, already public, already quotable. Each becomes a `Praise/` note with `public: true` and `solicited: true`.
2. **Experience section** (`/in/<handle>/details/experience/`). Fills the spine. Treat their own summary text as claims to verify, not facts, since most people last edited it years ago in a hurry.
3. **Their posts and the comments on them** (`/in/<handle>/recent-activity/all/`). Two things here: what they choose to talk about publicly, which is a read on positioning, and what people say back. Comments from named practitioners are underrated evidence.
4. **Skills endorsements.** Weak evidence, but the ranking shows how the market perceives them, which is worth comparing against how they see themselves. Record the gap, do not treat the endorsements as skill evidence.
5. **Messages** (`/messaging/`), only if they say yes. Inbound recruiter interest shows what the market already thinks they are. Praise in DMs is often the most candid there is. This is private material: quotes go in `Praise/` with `public: false`, raw threads stay in `Sources/`.

If Chrome is unavailable, ask them to copy their profile text and paste it, and to paste their recommendations. Most of the value survives.

## granola

The best source for unsolicited praise, because nobody performs in a working meeting.

Use `list_meetings` or `query_granola_meetings` to survey, then `get_meeting_transcript` for the ones worth reading. Do not pull every transcript in the account on the first run. Start with the last six months, report what it yielded, and offer to go further back.

Prioritise: retrospectives, one-on-ones with managers, performance conversations, project kickoffs and wrap-ups, and any meeting where they presented.

Look for four things:

- **Praise**, especially the offhand kind. "That thing you did with the docs saved us" in the middle of a status call is worth more than any recommendation, because nobody says it for effect.
- **Credit given elsewhere.** People describing their work to third parties, in their absence or presence, is external validation of scope.
- **What they are asked for.** Whoever gets pulled into a conversation for a particular reason, repeatedly, has a distinctive skill they probably do not list. Count the pattern across meetings, not the individual instance.
- **Scope evidence.** Decisions they made, things escalated to them, what they were held accountable for.

Every extraction gets `who`, `their_role`, `when`, `source: granola`, `public: false`, `solicited: false`.

Be careful with the difference between politeness and praise. "Thanks so much, this is great" in a handover is filler. Specific praise that names an effect is evidence. Record the second kind and let the first go.

## gmail

Search their mail for praise threads and review material. Useful queries: thank you, great work, congratulations, nice job, performance review, feedback, promotion, plus recommendation requests they have sent or received. Also search for their own name in third-party contexts, which surfaces people introducing them to others - a rich and completely invisible source of positioning language.

Customer and community email is high value in developer relations, since it names the person who was actually helped.

## calendar

Not evidence in itself, but the fastest read on where their time actually went and who they really worked with. Look at recurring meetings, who appears most often, and how the mix shifted over time. Use it to populate `People/` and to notice roles or projects nowhere in the record. Then feed the surprises into the interview queue.

## github

Only if they write code. Contribution history, repos owned, notable pull requests, and review comments received. For developer relations, sample apps, SDK work, and documentation commits matter more than raw contribution counts. A well-received PR review comment is praise.

## drive

Search for resumes, review documents, strategy docs they authored, and presentations. Documents they wrote are evidence of scope and also voice samples. Treat anything found here the same as `Incoming/`.

## slack

Their own saved items, DMs, and channel mentions. Praise in a public channel is quotable to them even if not publicly; praise in a DM is candid. Same handling as Granola.

## After any run

1. Write the notes, using the templates in `<vault>/Templates/`. Link them: wins to roles, praise to wins, people to roles.
2. Update the coverage table and ingestion log in `STATE.md`.
3. **Add open questions from what you found.** This is the part that makes ingestion compound. A win with no numbers, a role with no ending, a piece of praise about something not otherwise recorded, a repeated pattern nobody has named - each becomes a ranked question with a reason.
4. Commit.
5. Report: what was found, the most surprising thing in it, and what to run next. If ingestion surfaced something that contradicts the current `PROFILE.md`, say so directly.

From bendechrai/career-brain, MIT.

More from bendechrai/career-brain