lead-list-crm

Lead List → CRM

End-to-end recipe to (1) build a contact list, (2) enrich it with emails + LinkedIn member ids, (3) send paced LinkedIn connection requests, and (4) manage it in a tiny local web CRM with two buttons. Each stage writes a CSV; the CSV is the only state. Keep everything in one working folder.

Assets in this skill:

• assets/crm_app.py — the two-button web CRM (Send invites N · Update CRM). Requires Unipile.
• assets/list_crm.py — standalone list CRM (zero deps, click-to-cycle status). No Unipile needed.
• assets/prepare_crm.py — converts any contacts CSV into the CRM schema.

Pick the tools per stage; you don't need all of them. Crustdata alone can do build+enrich; Deepline is great for CSV/play workflows; Unipile drives LinkedIn.

---

Prereqs

• Crustdata MCP (company/people search + enrich). Check balance first: crustdata_credits_check (free).
• Unipile for LinkedIn actions. Creds live in ~/.config/unipile/.env:

  UNIPILE_API_KEY=...
  UNIPILE_DSN=https://apiNN.unipile.com:PORT
  

  Load them before any unipile/script call: set -a; . ~/.config/unipile/.env; set +a, then verify unipile accounts list. A 401 missing_credentials means the token/DSN is wrong — get a fresh token + matching DSN from dashboard.unipile.com (and confirm a LinkedIn account is actually connected).
• Deepline (optional). deepline auth status; for CSV viewing use deepline csv render. Note the legacy Python CLI ≠ the SDK CLI — the deepline-plays skill targets the SDK one.

---

Stage 1 — Build the list

Goal: a CSV with at least name, company. Two good paths:

Crustdata (fastest for ICP search):

1. crustdata_company_search_db — filter by industry, headcount, location, funding → company set.
2. crustdata_company_identify (free, batches up to 25 names comma-separated) → resolve company ids/domains.
3. crustdata_people_search_db (3 credits/100) — filter current_employers.name + current_employers.title + current_employers.seniority_level to get the role-holders. Returns linkedin_profile_url (often the hashed /in/ACoAA… form, which still resolves).
   • Title fields are a closed vocabulary for =/in; use [.] substring for free text, and crustdata_autocomplete_person to find exact values.
   • For a leadership sweep: title OR of ["CEO","Founder","Chief","Head of","VP","Director"] + seniority in ["CXO","Vice President","Director","Owner / Partner"].

Deepline (CSV-native / plays): use the deepline-gtm or deepline-plays skills to search companies/people or process an input CSV. Export the run to a CSV.

For a big fan-out (many companies × many titles), spawn one search agent per company in parallel (Agent tool or a Workflow) and concatenate the results — far faster than looping.

Output of this stage: contacts.csv (name, company, role, linkedin_url).

Stage 2 — Enrich

Emails (Crustdata):

• crustdata_people_enrich with linkedin_profile_url (comma-separate up to 25 per call) + include_business_email: true (+1 credit each). Returns business_email and the vanity linkedin_flagship_url.
• include_personal_contact_info: true (+2 credits) adds personal email/phone but is plan-gated — it 403s on many accounts and still burns credits on the failed call. Test the flags on ONE profile before running the batch.
• For 26+ profiles / max fill rate, crustdata_batch_people_enrich exists but is also access-gated (403 → ping [email protected]).

LinkedIn handle + connection degree (Unipile): for each contact, GET /api/v1/users/{member_id}?account_id=… returns public_identifier (→ clean linkedin.com/in/<slug>), headline, location, and network_distance (FIRST/SECOND/THIRD_DEGREE). member_id = the part after /in/ in the Crustdata URL. Pace ~1s/call; this is how you learn who's already 1st-degree vs reachable.

Stage 3 — Prepare the CRM file

Convert whatever you built into the CRM schema:

python <skill>/assets/prepare_crm.py contacts.csv crm.csv

It maps common header aliases and extracts member_id from the LinkedIn URL. Required columns it produces: person_name, company, role, member_id, linkedin, invited, invited_date, invite_status, messaged, replied, last_touch, next_followup, notes. Rows without a member_id stay in the list but can't be invited until you resolve one.

Stage 4 — Run & manage the CRM

set -a; . ~/.config/unipile/.env; set +a
nohup python <skill>/assets/crm_app.py > crm_app.log 2>&1 &
# open http://127.0.0.1:8787

Env knobs: CRM_CSV (default crm.csv), CRM_PORT (8787), CRM_TZ (default America/Los_Angeles), UNIPILE_ACCOUNT_ID (else auto-detects the first connected LinkedIn account).

Button 1 — Send invites N: sends connection requests to the next N uninvited rows (those with a member_id), via POST /api/v1/users/invite ({provider_id, account_id}). Paced with random 35–75s gaps + a 150–240s cooldown every 10 to avoid tripping LinkedIn's automation alarms. Runs in a background thread; marks invited=yes + date as it goes.

Button 2 — Update CRM: re-checks each invited row — network_distance (accept → invite_status=accepted), scans /api/v1/chats (attendee_provider_id) to flag who you messaged, reads chat messages (is_sender==0 ⇒ inbound) to flag replied, then recomputes next_followup:

• replied → today (HOT, respond now)
• accepted, not messaged → today (send first touch)
• accepted, messaged, no reply → +3 days
• unaccepted invite → no follow-up (just keeps getting re-checked for accepts)

Dates display human-readable (Wed, 17 Jun) in the chosen timezone; ISO is kept under the hood for sorting.

Follow-up reminders

The CRM's next_followup column is the durable source of truth. For an active nudge, schedule a one-shot reminder (CronCreate, recurring:false, durable:true) a few days out whose prompt = "load unipile env, cd to the folder, re-run the CRM update, report new accepts / replies (HOT) / due follow-ups, ask before sending." In-session crons die when the session ends — for a hard backup, add a calendar event.

---

---

Standalone list CRM (no Unipile required)

Use assets/list_crm.py when you have a contacts CSV and just need to triage, track outreach status, star, and take notes — no LinkedIn automation yet.

CRM_SRC=contacts.csv \
CRM_CSV=crm.csv \
CRM_TITLE="My List" \
CRM_CAT_COL=category \
python <skill>/assets/list_crm.py
# open http://127.0.0.1:8899

Env knobs: CRM_SRC (source CSV, read once to bootstrap), CRM_CSV (working file, default crm.csv), CRM_PORT (8899), CRM_TITLE, CRM_CAT_COL (column used as category filter pill), CRM_STATUSES (JSON array to override defaults).

How it works:

• On first run, copies source CSV → CRM CSV, adding status / priority / notes / last_touch columns.
• Serves a /data JSON endpoint + a single-page client. All columns from the source CSV render as table columns automatically.
• URL and http:// values auto-render as links; the CRM_CAT_COL column renders as a pill with a dropdown filter.

Click-to-cycle status state machine. Each row has a colored pill for status. One click advances to the next state; wraps around at the end. No dropdown, no form submit — just click. This is far faster to use than a select when triaging 100+ rows.

Default funnel (covers both LinkedIn and email outreach paths):

To review → Following → Sent invite → Connected →
Sent message → Sent email → Replied → Talked → Pass → (wraps)

Override via CRM_STATUSES='["Cold","Warm","Hot","Pass"]' for a simpler funnel.

Status pill colours are defined in STATUS_CSS at the top of list_crm.py. Add a new line per status — the data-v attribute on the pill drives CSS targeting.

Atomic CSV writes. All saves go through tmp + os.replace() — no truncation risk if the process dies mid-write.

Graduating to the Unipile CRM. Once you've triaged the list and want to start sending LinkedIn invites, run prepare_crm.py on your CRM CSV to convert it to the crm_app.py schema (which adds member_id, invited, invite_status, etc.).

---

Gotchas learned the hard way

• Deepline CSV viewer caches by file path. After editing a CSV, deepline csv render/csv show keep serving the old version (and restarting the renderer can flush its cached copy back over your edits). Fix: render under a fresh filename, and stop the renderer before editing a file it holds.
• Crustdata charges for failed enrich calls that include a gated flag (include_personal_contact_info). Test flags on one profile first.
• crustdata_people_search_db returns 0 silently when a closed-vocab value is slightly wrong ("VP" vs "Vice President"). Use autocomplete or [.] substring.
• Hashed /in/ACoAA… URLs work for invites and users lookups; they resolve to the real profile and Unipile returns the vanity slug on enrich.
• LinkedIn invite ceiling is ~100–200/week. 50 paced in a run is assertive; let a big batch breathe a few days before the next.
• already_invited_recently (422) on invite just means a pending invite already exists — count it as skipped, not failed.
• Accepts trickle in over 3–7 days — re-run Update over several days, don't judge accept rate on day one.