draft-outreach

---
name: draft-outreach
description: Drafts LinkedIn connection notes, cold emails, referral requests, cover letters, follow-ups, and thank-yous for a DS/CS student's internship, co-op, or new-grad SWE / data science / ML search, fact-checked line by line; connection notes go to a queue the student approves, everything else stays a draft the student sends. Use when the user asks to write outreach, a connection note, a cover letter, a referral ask, a thank-you, or a follow-up.
---
# draft-outreach

## Inputs
- Recipient: a `contact_id` from `profile/contacts.md`, or a person the user names.
- Role: a `role_id` from `profile/roles.md` (JD + `### Fit`), if any.
- `profile/facts.json` (the only source of claims about the user).

## Draft types
| Type | Length | Goal |
|---|---|---|
| LinkedIn connection note | ≤ 200 characters | accept the invite; one tie + one fact |
| Cold email to manager/engineer | 90–130 words | 15-min chat or referral consideration |
| Alumni / referral ask | 80–120 words | referral; lead with the shared school/club/employer |
| Cover letter | ≤ 250 words, 3 paragraphs | why them, why me (2 facts), close |
| Follow-up | 40–70 words | after `follow_up_after_days`, with something new |
| Thank-you (post-interview) | 60–100 words | one specific detail from the conversation |

## Steps
1. Pick the type; confirm the recipient's name, title, and source URL from `contacts.md`.
2. **Hook:** one specific, verifiable thing about *them* (their post, repo, talk, launch, or the shared tie exactly as `contacts.md` records it, e.g. "fellow UofT CS student").
3. **Value line:** 1–2 facts matching the role.
4. **One small ask.** Link resume/GitHub; no attachments unless asked.
5. Write subject (≤ 8 words, emails only) and body. Plain text, no emojis, no flattery.
6. **Fact check** every sentence about the user against facts.json and print the table:
   `sentence → fact id: "exact supporting string"`. Check specifically: current vs former (school/employer `current` flag), metric name + unit + value, degree name, titles, years (computed from dates). A sentence without support is rewritten or removed.
7. **Save:**
   - Connection note → append a `draft` row to `profile/connect-queue.md` (format in `linkedin-browser`) with the next `queue_id`, recipient URL, `sender_url` = `prefs.linkedin_profile_url`, and the exact note.
   - Everything else → `output/drafts/<date>-<role_id|contact_id>-<type>.md` with header To / Source URL / Subject / Body / Facts used.
8. **Stop for the human.** Show the drafts and ask: "Approve, edit, or skip?" For connection notes, approval means the user names the `queue_id`s ("approve q3, q5"); `linkedin-browser` `connect` mode then sends exactly those. Editing a note resets its row to `draft`.
9. When the user says they sent something, or `linkedin-browser` reports `sent`, hand it to `track-pipeline`.

## Rules
- The human sends every email, DM, and InMail. Connection requests go out only through `linkedin-browser` connect mode, only for approved rows.
- No invented facts, mutual connections, or prior conversations.
- Respect `outreach.max_new_per_week` for emails/messages; invites are capped in `linkedin-browser`. One follow-up per thread unless the user asks.