Two-phase LinkedIn lead pipeline driven by linkedin-cli. Phase A imports leads from a search URL or filters, qualifies them against a configurable ICP via sub-agent, and stores them in a local SQLite database with round-robin assignment across one or more LinkedIn accounts. Phase B runs on a schedule per account — sends connection invites up to a daily limit and withdraws stale pending requests. Use when the user wants to grow their network from LinkedIn searches, manage outgoing invites at scale, ask status questions (counts, conversion, pending older than N days, last imports), pause/resume an account, change ICP, or install the recurring scheduler.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Two-phase LinkedIn lead pipeline driven by linkedin-cli. Phase A imports leads from a search URL or filters, qualifies them against a configurable ICP via sub-agent, and stores them in a local SQLite database with round-robin assignment across one or more LinkedIn accounts. Phase B runs on a schedule per account — sends connection invites up to a daily limit and withdraws stale pending requests. Use when the user wants to grow their network from LinkedIn searches, manage outgoing invites at scale, ask status questions (counts, conversion, pending older than N days, last imports), pause/resume an account, change ICP, or install the recurring scheduler.
This skill turns a Sales Navigator (or regular) search into a managed pipeline:
search → qualify (you, via sub-agent) → store → invite on schedule → check pending → withdraw stale.
All state lives in a local SQLite database. Every LinkedIn action goes through
linkedin-cli (the linkedin binary). You orchestrate via the Node scripts
under scripts/.
Vocabulary
Term
Meaning
Account
A LinkedIn account registered in linkedin-cli AND added to this skill's DB. The DB stores per-account policy (daily_invite_limit, active_start/active_end hours, max_pending_days, paused).
Lead
A qualified person row in leads. PK is hashed_url (Sales Nav hashed URL for nv imports, or publicUrl for st imports). Has exactly one owner_account.
List
A free-text list_name attached at import time (e.g. "VP of Sales TOP 100"). Used for filtering and conversion analytics.
Batch
A row in import_batches. Created by import.mjs prepare; transitions `pending_qualification → committed
Status
A lead's lifecycle state: `not_connected → pending → (connected
Retry policy
Global max_connect_attempts setting. When an account's attempt fails (we withdrew a stale pending, OR the person declined/expired), the lead is reassigned to another untried account and set back to not_connected — until max_connect_attempts distinct accounts have tried, then it becomes exhausted. 1 = no retry (default); all = try every active account.
Phase A
Import — runs only when the user triggers it. Includes the LLM step (you, via sub-agent).
Phase B
Network maintenance — network-invite + network-pending. Runs on a schedule. NEVER calls an LLM.
First-run setup
1. Verify Node ≥ 20:node --version.
If missing — print the OS-specific install command and stop:
If the output is Cannot find module 'better-sqlite3':
npm install --omit=dev
Then re-run doctor. (Alternative: node scripts/doctor.mjs --fix does this automatically.)
3. For each FAIL in the doctor output, apply the remediation:
Check name
Remediation
linkedin-cli
npm install -g @linkedapi/linkedin-cli
cli-accounts
Ask the user for their Linked API Token and Identification Token (link: https://app.linkedapi.io), then linkedin setup --linked-api-token=<a> --identification-token=<b>. Repeat per LinkedIn account they want connected.
db
Auto-fixed by any script on first invocation, or explicitly: node scripts/db.mjs init
db-accounts
Run linkedin account list (prints a table; the * marks the active account) and register each one here: node scripts/account.mjs add --name <short-name> --cli-account "<exact name from linkedin account list>". The short name is what every other command takes; the cli-account is the mapping.
scheduler
Should pass automatically. On headless Linux without systemd-user, doctor falls back to cron.
4. Re-run node scripts/doctor.mjs --json until "ok": true.
5. Set the connection pace — ask once, apply to all accounts. Ask the user a single
question (not per account): "By default each account sends at most one connection request
every 15 minutes — keep 15, or change it?". Apply their answer to every account via
--min-invite-interval <N> (either pass it on each account.mjs add, or
account.mjs update --name <acct> --min-invite-interval <N> for all afterward). Default is
15. Let the user know they can fine-tune it per account later just by asking (e.g. "make
kiril one every 30 minutes") — it is a per-account setting, this question just sets a common
value for everyone.
6. Set the retry policy. Ask the user: "If someone doesn't accept the request, should
we try connecting from another account? (no / a specific number of accounts / all of them)".
Then:
node scripts/settings.mjs set max_connect_attempts 1 # no retry (default)
node scripts/settings.mjs set max_connect_attempts 2 # original + 1 more
node scripts/settings.mjs set max_connect_attempts all # every account
7. Enable the background scheduler (only after at least one account is registered):
node scripts/schedule.mjs install
This installs one platform-native background task that keeps the pipeline running
on its own. When talking to the user, describe it as "the pipeline now runs in the
background and sends invites during each account's active hours" — do not expose the
scheduler's internal wake-up frequency (the tick) or other plumbing. (The invite pace
from step 5 — "one connect every N minutes" — is a real user-facing setting and fine to
discuss; it's the tick's 5-minute heartbeat that stays hidden.) See the Phase B and
Scheduler sections below for how it actually works.
8. Tell the user the next step and offer to do it. Setup alone sends nothing — the
pipeline is empty until leads are imported. End onboarding with a concrete call to action,
e.g.: "You're all set. To start, give me a LinkedIn or Sales Navigator search URL (or
search filters) and a name for the list, and I'll import and qualify your first batch of
leads." If the user provides one, proceed straight into Phase A below. Do not end the
setup conversation without this prompt.
Phase A — Importing leads (interactive)
Triggered by the user via wording like:
"import leads from this search ..."
"add leads from this Sales Navigator URL"
"add a list called 'X' from this search"
Step 1 — Prepare
Always ask the user for a limit first. Before running prepare, ask "how many of the
found leads should I take?" — the user gives a number, or says "max" for the maximum. The
maximum depends on the search type (these are the Linked API / LinkedIn caps):
Sales Navigator (nv): 2500
standard search (st): 1000
--limit is required; prepare errors if it is missing. Pass --limit max (or all)
for the cap, or a number (clamped to the cap, with limit_capped_to_max: true reported).
Auto-detect: if the URL contains /sales/, pass --type nv; otherwise --type st.
Either --search-url or filter flags must be provided.
The script:
Runs the LinkedIn search via linkedin-cli (workflow run for URL-based, native CLI for filters)
Normalizes results
Dedupes against existing rows in leads (skipped count is reported)
Writes the new candidates to <data_dir>/tmp/qualify-<batch-id>.candidates.json
Creates an import_batches row in state pending_qualification
Returns the batch id + the candidate file path + the expected result file path + the path to the qualification prompt
Step 2 — Qualify (YOU, against the user's ICP)
Qualification is filtering candidates against the user's ICP (Ideal Customer Profile) —
their definition of who is a good lead and who to filter out. The ICP is user-owned and
must come from the user, never from a hardcoded list. It is stored in the
icp_definition setting and persists across imports.
a. Make sure there is an ICP.prepare's output includes icp_configured and the
current icp_definition.
If icp_configured is false: interview the user before qualifying. Ask concrete
questions — which roles/seniority to target, which industries/company types fit, company
size/stage if relevant, locations to include or exclude, and any hard exclusions
(competitors, students, specific titles). Summarize what you heard back to them, then save
it straight into the database via stdin (no stray files):
If icp_configured is true: show the user the current ICP in plain language and ask
whether to use it as-is or tweak it for this list. If they tweak it, re-save it the same way.
b. Qualify each candidate. Read the candidate file (JSON array of
{hashed_url, public_url, full_name, position, location}) and the qualification contract at
config/qualification-prompt.md. Judge every candidate against the ICP. For more than ~25
candidates, chunk the work and delegate each chunk to a sub-agent (Task tool in Claude Code,
or the equivalent in other hosts), passing the ICP + the contract + the chunk. Each must
return [{hashed_url, suitable, reasoning}] covering EVERY lead, preserving hashed_url,
where reasoning cites the actual ICP criterion that drove the decision.
Use a cheap, fast model for the qualification sub-agents. This is a bounded
classification task (role + location vs ICP → boolean + one-line reason), not deep
reasoning — so the heaviest model is a waste of money at lead volume. In Claude Code, spawn
the qualification sub-agents with model: "haiku" (the Task tool's model parameter); in
other hosts pick their equivalent small/fast model. Keep the orchestration, the ICP
interview, and the final report on the main model — only the per-chunk classification goes to
the cheap tier. If the ICP is unusually nuanced and you see many borderline calls, raise the
tier for that import. The stored per-lead reasoning lets you spot-check cheaply.
Concatenate all results and write them to the expected result file path from prepare.
(Non-agentic context: the user can write the result file by hand or with any model; nothing
in the skill enforces a specific provider.)
For each suitable: true: round-robin assigns owner_account from active (non-paused) accounts in alphabetical order, starting after the last-assigned account (cursor persists across imports). Inserts the lead with status='not_connected'.
After committing, report the decision transparently so the user understands the filter:
state how many were kept vs filtered, and give a few concrete sample reasons from both sides
(e.g. "kept: Head of Sales at a B2B SaaS — matches target role; filtered: Software Engineer —
not a targeted role"). The per-lead reasoning is stored on each lead and is also queryable
later via node scripts/lead.mjs show <id> or query.mjs.
Other batch commands
node scripts/import.mjs list [--state pending_qualification|committed|aborted]
node scripts/import.mjs show --batch <id>
node scripts/import.mjs abort --batch <id> # cancel a pending_qualification batch
Phase B — Network maintenance (scheduled, distributed)
Phase B is not a single daily batch. The background scheduler does small,
resumable units of work spread across each account's active hours. Invites and
pending checks are decoupled — they run on their own cadence.
On each wake-up, for every active account that is within its active window
(active_start–active_end, local time):
Invites (write, rate-sensitive): send one invite if both
the daily quota (daily_invite_limit) is not yet reached, and
at least min_invite_interval_minutes have passed since the last invite.
min_invite_interval_minutes is the explicit "no more than one connect every N
minutes" control (default 15). The effective daily ceiling is the tighter of the
daily limit and what the interval allows inside the window.
Pending checks (mostly reads, low-risk): process up to pending_batch_size
due pending leads (status check, and withdraw if still pending past
max_pending_days). This runs independently of the invite decision and is
not throttled by the invite interval, so a backlog of stale pending requests
drains quickly instead of one-per-wake-up.
Both can happen in the same wake-up.
Why this shape matters (and what to tell the user if they ask):
Each LinkedIn operation is written to the DB immediately. If the machine sleeps or
a run is killed mid-operation, the next wake-up just continues from the current DB
state — there is no batch to resume and nothing to roll back.
Daily quota is recomputed from the runs table every time (bounded to the local
calendar day), so it stays correct across interruptions and restarts.
Invites are paced by an explicit interval; pending checks are not — a read is cheap,
a write is rate-limited.
You normally never run Phase B by hand. For testing or a deliberate one-off "drain
now" (ignores pacing, respects the daily quota and active-window checks inside the
scripts only loosely — use with care):
node scripts/network-invite.mjs --account <name> --limit 1 # exactly one invite
node scripts/network-pending.mjs --account <name> --limit 1 # exactly one pending check
node scripts/network-run.mjs --account <name> # full invite + pending sweep now
Invite outcomes
For each not_connected lead within the day's remaining budget, runs the workflow:
data.then.error.type includes alreadyPending → status='pending'
data.then.error.type includes alreadyConnected → status='connected'
data.then.error.type signals an account-level limit on the action category (e.g.
limitExceeded, a rate limit) → the lead stays not_connected (NOT a per-lead error) and
this account's run backs off (stops for the cycle); the lead is retried on a later wake-up.
data.then.error.type is noteLimitExceeded → the account has reached LinkedIn's
personalized invitation-note limit. Treat it as account-level gating, not a per-lead error:
leave the lead not_connected, back off, and retry later or send future invites without notes.
data.then.error.type is requestNotAllowed ("LinkedIn has restricted sending a connection
request") → ambiguous, disambiguated by pattern: a streak (2+ in a row with no successful
invite between) = the account's weekly invite limit → leave not_connected and back off (not
the lead's fault); an isolated hit = the person restricts invites → counted against the
lead, and after RESTRICTED_LEAD_ATTEMPTS (2) isolated hits the lead is closed as exhausted
so it never hangs. Never burns a whole queue on a weekly-limit burst.
linkedin-cli exit code 3 (subscription or plan required), 4 (account issue), or 6 (rate limit)
aborts the whole run immediately — no further leads are touched. Other non-zero exits mark the
lead as error and continue.
Pending outcomes (with cross-account retry)
For each pending lead where sent_at is older than max_pending_days:
pending (still) → linkedin connection withdraw <public_url> → on success this is a
failed attempt → apply retry policy
other → leave as pending, log run as error
Retry policy (resolveFailedAttempt). On a failed attempt, look up how many distinct
accounts have already invited this lead (from the runs table). If that count is below
max_connect_attempts (global setting) AND there is an active account that has NOT tried
this lead yet, reassign the lead to the least-loaded such account and set it back to
not_connected (it re-enters the invite flow under the new account). Otherwise mark it
exhausted (terminal). With the default max_connect_attempts = 1, every failed attempt
goes straight to exhausted (no retry).
Answering arbitrary status questions
The user will ask things like "how many pending on ?", "imports last 7 days?",
"which lists convert best?", "why is lead X in error?". Substitute the user's real
account/list names. Use this decision tree:
All timestamps are SQLite datetime('now') strings in UTC.
Retry policy (global settings)
node scripts/settings.mjs list
node scripts/settings.mjs get max_connect_attempts
node scripts/settings.mjs set max_connect_attempts 2 # try original + 1 more account
node scripts/settings.mjs set max_connect_attempts all # try every active account
node scripts/settings.mjs set max_connect_attempts 1 # no retry (default)
max_connect_attempts is the number of DISTINCT accounts that may attempt one lead. When
a request goes unaccepted (withdrawn stale pending, or declined/expired), the lead is
handed to the least-loaded untried account until this many accounts have tried, then it is
exhausted. Frame it to the user as "if someone doesn't accept, try from N other accounts".
Account management
<account> below is a placeholder — the user picks their own short name; it maps to
a real linkedin-cli account name (from linkedin account list). The skill ships
with no accounts and no predefined names.
active_start/active_end — when (daily hours, local time) invites go out.
min_invite_interval_minutes — how fast (minimum gap between two invites,
default 15). This is the direct "no more than one connect every N minutes" knob.
daily_invite_limit — how many per day (hard cap).
status.mjs reports effective_max_per_day = the tighter of the daily limit and what
the interval allows inside the window. pending_batch_size controls how many stale
pending requests are checked per wake-up (independent of invites). When discussing with
the user, frame these as plain-language behavior ("invites go out 9am–6pm, at most one
every 15 minutes, up to 35 a day") — never in terms of the scheduler's wake-up frequency.
ICP and the qualification contract
Two separate things:
The ICP (who to keep / filter) is user-owned data. It lives in the local database
(the settings table, key icp_definition) — NOT in any file. View it with
node scripts/settings.mjs get icp_definition. Change it by piping the text via stdin:
node scripts/settings.mjs set icp_definition --stdin <<'ICP'
<the ICP text, multi-line, as many lines as needed>
ICP
(For a short one-liner, settings.mjs set icp_definition '<text>' also works.) Do NOT write
the ICP to a stray file in the repo or some tmp folder and load it from there — the file is
not where it lives, and it litters the workspace. Capture the ICP by asking the user; never
hardcode one. Changes take effect on the next import.
The qualification contract at config/qualification-prompt.md is the product-agnostic
scaffolding (how to judge + the JSON output format). You normally don't change it; it
references the user's ICP rather than containing one.
Scheduler
The scheduler is one OS-native background task (launchd / systemd-user / cron /
schtasks depending on platform) that wakes the pipeline periodically. The wake-up
frequency is an internal detail — the user-facing behavior is set by each account's
active hours and daily limit. Do not surface intervals to the user.
node scripts/schedule.mjs detect # reports launchd | systemd-user | cron | schtasks
node scripts/schedule.mjs status
node scripts/schedule.mjs install [--interval-minutes 5] # interval is internal; default is fine
node scripts/schedule.mjs uninstall
To inspect what the background runs are doing, read <data_dir>/logs/<account>-<YYYY-MM-DD>.log.
Common pitfalls (read before acting)
Two LinkedIn accounts in the same DB with the same cli_account mapping — undefined behavior; reject if the user tries it. Use account.mjs list to verify.
Lead PK across search types — Sales Nav (nv) returns hashed URLs; regular (st) returns public URLs. The same person from both search types becomes two rows. If users mix, mention it explicitly.
Interrupted mid-operation — each scheduler wake-up does at most one invite/check, persisted immediately. A sleep/kill loses at most that one in-flight operation; the next wake-up continues from DB state. There is no batch to resume. Leads not yet processed stay not_connected/pending and are picked up later.
The connected status on invite — only set when LinkedIn reports alreadyConnected at invite time (the person was already a 1st-degree connection). Real new connections appear via network-pending (where connection status returns connected).
error status is terminal until reset — lead.mjs reset <hashed-url> moves it back to not_connected. The auto-pipeline does not retry errored leads on its own.
Renaming an account — leads' owner_account is ON UPDATE CASCADE; rename is safe. remove without --force refuses when leads exist; with --force they are orphaned (status queries will still include them but no scheduled run touches them).
schedule.mjs install without registered accounts — harmless, but nothing happens until accounts exist. Install order: doctor → add accounts → install scheduler.
Daily quota boundary — the per-day invite count resets at local midnight, not UTC. Quotas and sent_today are computed against the local calendar day.
Idempotency
db.mjs init, doctor.mjs, schema.mjs, status.mjs, lead.mjs show/list, query.mjs, schedule.mjs status — all safe to call multiple times.
import.mjs prepare creates a new batch every time — call once per intended import.
import.mjs commit refuses to run twice on the same batch.
schedule.mjs install overwrites any existing installation of the same service id.
Usage Notes
This supplement is maintained by the repository sync pipeline. It keeps the
imported upstream skill usable inside this curated collection when the upstream
source is intentionally concise.
Common Patterns
1. Confirm that the user's task matches the skill trigger.
2. Read the relevant project files or user-provided context before acting.
3. Choose the smallest reversible action that advances the task.
4. Run the verification command or manual check that proves the result.
5. Report the outcome, evidence, and any remaining risk.
Boundaries
Prefer the upstream workflow for Linkedin Growth; this section only adds local quality
guardrails.
Do not invent project facts when required files, vaults, services, or tools are
unavailable.
Stop and ask for clarification when the next action could overwrite user work,
expose private data, or change production state.
