LinkedIn outbound

One connected LinkedIn account, a target search, a daily budget of invites and messages, and a human who approves every batch of copy before it goes out. The agent does the search, the reading and the drafting; the invites and messages go out only after the gate.

Before you start

• Load the core jell skill (https://jell.run/SKILL.md) if it is not loaded. Confirm access with the free balance tool or jell balance.
• Run the free linkedin.accounts first. It lists the accounts the workspace already connected, which profile each one is, and the account_id to pass; ignore rows whose status is EXPIRED or PENDING (a sign-in link that was never finished). If none is connected, run linkedin.account with a name: it returns a hosted sign-in link the user opens once. Never ask for the user's LinkedIn password.
• Ask for the target (title, company type, location), the offer in one sentence, and the daily caps if the user has them. Default: 20 actions a day per account, invites and messages together, never both to the same person on the same day. Keep the log at linkedin/log.csv (columns: date, action, profile_url, provider_id, run_id, note_or_message, status) and the suppression list at linkedin/suppressed.txt, relative to the folder the user names; ask once.
• Inspect linkedin.search, linkedin.profile, linkedin.invite, linkedin.message once and show the prices. Quote the day: N searches, N profiles, N invites, N messages.

Steps

1. Search

jell run -c linkedin.search -i '{"keywords":"head of growth fintech","location":["London"],"network_distance":[2,3],"account_id":"<id>","limit":10}' --max-cost 0.05 --wait 60 -o search.json

A search costs a fraction of a cent. location, company and industry take place or company names as well as LinkedIn ids; a name resolves to its first LinkedIn match ("London" becomes "London Area, United Kingdom"), so use the name LinkedIn shows for the place you mean. A name that matches nothing is dropped, the search still runs, and the result carries dropped_filters and a note: check for them before trusting the list, and fall back to a broader place name, the id, or the city in keywords. Classic search returns a page of 10 whatever the limit and bills the page; use cursor for the next one. Rows carry provider_id, public_identifier, profile_url, name, headline, location, degree: the company is in the headline, and location comes back in the account's own LinkedIn language ("Londres, Angleterre, Royaume-Uni" on a French account), so never filter rows on an English place string. Expect noise (a "Head of People" for a growth search, an anonymous "LinkedIn Member" row): the profile read in step 2 is what qualifies a row, not the search. Keep name, headline, profile URL. Before anything else run the free dedupe pass: jell history --file urls.txt with the profile URLs and read the contacted column, which is who this workspace already invited or messaged; drop them. seen is always yes right after a search and means nothing here.

2. Read before writing

jell run -c linkedin.profile -i '{"profile_url":"https://www.linkedin.com/in/example","full":true,"account_id":"<id>"}' --max-cost 0.10 --wait 60

The profile read returns provider_id, which linkedin.invite and linkedin.message accept in place of the URL. Without full: true it carries only the name, headline, location, degree and counts; full adds the summary, the work history and the skills (LinkedIn throttles it, so read at most the people you will write to). There is no posts field. Pull the one detail the note will use: a recent role change, a company event, a line from the summary. A note without a real detail is a template; do not send templates.

3. Draft the notes

One per person, under 300 characters (LinkedIn's cap; shorter reads better), no link, no pitch, one reason the connection makes sense. The offer sentence belongs in the step 5 message, not in the note. Put them in a table: name, profile URL, the detail used, the note.

Gate 1: show the table and wait. The user edits or approves. Nothing is sent before this.

4. Invite

jell run -c linkedin.invite -i '{"profile_url":"https://www.linkedin.com/in/example","message":"<approved note>","account_id":"<id>"}' --max-cost 0.05 --wait 60

Before the batch, run the free headroom check:

jell run -c linkedin.invitations_sent -i '{"account_id":"<id>"}' --wait 60

It counts the account's pending invitations over the last 24 hours and 7 days and returns remaining_today, the most to send now (LinkedIn allows roughly 100 a day and 200 a week). Send no more than the lower of remaining_today and the daily invite cap; when it is 0, send nothing and say so. The counts are a floor (accepted and withdrawn invitations leave the pending list), so the user's own cap still rules. Stop the batch at the first provider_rate_limited; it bills nothing and is never retried.

Up to the daily invite cap. Log each one: profile URL, date, run ID, note.

5. Message after accept

On the next run, linkedin.messages with unread: true and after set to the last run's timestamp lists the conversations with something new, each with the person's name, profile_url, headline and degree; an empty list means nobody wrote, and is a success, not an error. Acceptance is not in the inbox: linkedin.profile on each invited person shows degree 1 (connected: true) once they accepted. For accepted connections that have not replied, draft one message each: the offer in two sentences, one question, no attachment.

Gate 2: show the messages and wait. Then:

jell run -c linkedin.message -i '{"profile_url":"https://www.linkedin.com/in/example","text":"<approved message>","account_id":"<id>"}' --max-cost 0.05 --wait 60

Up to the daily message cap. Use chat_id instead of profile_url when continuing an existing thread.

6. Triage replies

Read each conversation with chat_id: messages carry text, sent_by_me, sender_name, sender_profile_url and sent_at. For every reply: interested, not interested, wrong person, question, out of office. Draft a response for interested and question; the user sends it or approves it. Add not interested to linkedin/suppressed.txt. Never message someone who said no.

Rules

• Caps are per account per day and shared across invites and messages. Never exceed them, and start lower on a fresh account (5 to 10 a day for the first week).
• One note per person, written from that person's profile. No bulk template, no link in a connection note.
• The agent drafts; the user approves each batch. A scheduled run may search, read and draft on its own, and may send only copy the user approved in an earlier gate.
• Respect LinkedIn's terms and the user's: this skill sends through the user's own account on a hosted connection, at human pace, with a human in the loop. It never scrapes beyond what the capabilities return and never automates likes, comments or posts.
• max_cost on every run, --wait 60 on every example (a bare run returns a receipt, not a result). Every action logged with its run ID. Report the day's cost as the sum of every run's billing.charged, not a wallet difference: the balance is shared by every session on the key.

Output

A daily briefing: invites sent (with the log), acceptances since last run, messages sent, replies by class with the drafts to approve, and the running totals for the week. Plus the CSV log the next run reads.