Email Waterfall

Finds a verified work email for every row of a list by trying providers in order, cheapest first, and stopping at the first verified hit. It never builds an email from a name pattern, it never pays twice for the same person, and anything it cannot verify is kept with its status so it can go to a LinkedIn track instead of a live email sequence.

How to apply this skill

When asked to fill in or verify emails for a list:

1. Gather inputs: a CSV with first_name,last_name,company_domain,linkedin_url (extra columns are kept), the per-run spend cap in USD, and whether the GetLeads daily allowance covers the row count (GetLeads MCP get_fair_use, free).
2. Run scripts/outbound/waterfall.py from the scripts/ folder, dry run first, then for real in batches of 150. Do not hand-pick emails, do not reorder providers by habit, and do not add pattern guessing.
3. Report: the <out>.report.md numbers (% verified, % catch_all, % unknown, % not_found, hits per provider, spend), five sample rows per status, and which rows have domain_match=no. Only verified rows may go to Smartlead.

Do not: treat a catch-all or unknown address as sendable because the count looks small, or trust a verified email at a domain that is not the company on the list.

Setup

cd skills/email-waterfall/scripts
pip install httpx python-dotenv pyyaml        # or: uv run --with httpx --with python-dotenv --with pyyaml python -m outbound.waterfall ...
python -m outbound.common --selftest          # prints OK; no network
python -m outbound.waterfall --selftest       # prints OK; no network

Keys: GETLEADS_API_KEY in the repo .env (the scripts search upward for it). Deepline: deepline auth status must show a workspace; install from https://code.deepline.com/INSTALL.md. Credit prices are constants at the top of waterfall.py and common.py (DEEPLINE_CREDIT_USD, GETLEADS_CREDIT_USD); edit them to your plans so the cap is honest.

The waterfall

Default order, one line to change (PROVIDERS in waterfall.py, or --providers):

Verification: a provider that already returns valid/VERIFIED skips the verifier; everything else goes to MillionVerifier (0.06 credits), and a catch-all answer is re-checked once with BounceBan.

Stop rule per row, in this order:

verified   -> keep, stop
catch_all  -> keep with status, stop (a domain property: the next provider returns the same domain)
unknown    -> keep with status, stop
discard    -> drop the candidate (invalid, disposable, spam trap), try the next provider
exhausted  -> not_found

Output columns added to every input row: email, email_status, email_provider, verifier, mx_class, domain_match, miss_reason. Rows are never dropped; dedupe happens before any spend (LinkedIn URL first, then name + domain).

domain_match is yes when the email's domain is the company domain or shares its main label (clorox.com and thecloroxcompany.com match), no otherwise. A verifier proves the mailbox exists, not that it is your prospect; no rows are usually a namesake at another company or a person who has moved. Review them by hand. Observed on a real 10-row run: 2 of 9 hits before the explicit provider order fixed it, 1 of 10 after.

Steps

1. get_fair_use on the GetLeads MCP. Stop and say so if the allowance cannot cover the rows.
2. Plan: python -m outbound.waterfall ../../../leads.csv --out out/leads_enriched.csv --dry-run. A dry run writes only <out>.dryrun.csv and <out>.dryrun.report.md, never the real output, and spends nothing.
3. Test on 10 rows for real, read the output, then continue.
4. Real run: python -m outbound.waterfall <csv> --out out/leads_enriched.csv --batch 150 --cap-usd N. The cap is a total for the run; Deepline plays get a hard --max-credits derived from the remaining budget. A cap abort keeps every processed row and marks the rest cap_aborted. Re-runs replay cached calls at $0 (.cache/calls.jsonl, git-ignored, contains emails; delete it when a project ends).
5. Read out/leads_enriched.csv.report.md. If not_found is above 30%, say so and ask before adding deepline-play (Deepline's own prebuilt waterfall, which picks its own vendors) as a last leg of --providers.
6. Split: verified is the only set eligible for Smartlead. catch_all, unknown and not_found go to a LinkedIn-only track.
7. Show the table and wait for approval before any upload (/cold-email-starter-kit or /smartlead-campaign-upload-public).

Quick reference

Measured on 10 real rows (Oct 2026): 10/10 verified, 7 by LeadMagic, 2 by GetLeads, 1 by Prospeo, $0.33 total; the earlier Deepline-picks-vendors run found 9/10 for $1.22 with two namesakes. The hits per provider line is the data for reordering.

Common mistakes

• Guessing emails from a name pattern — the script refuses to; do not add it by hand.
• Reordering providers by habit instead of by the hits per provider line — the provider that carries your market should run first.
• Trusting domain_match=no rows — FullEnrich and Prospeo return a real mailbox for the person's current employer, which is not always the company on your list.
• Skipping dedupe, so the same person is paid for twice.
• Sending catch_all or unknown rows to Smartlead.
• Running the full list before looking at the first 10 rows.
• Setting --cap-usd from the estimate alone: real Deepline rows cost 1 to 5 credits each, and the hard cap now follows the remaining budget, so give the run the budget you actually mean.

Related skills

• /blitz-list-builder, /competitor-engagers, /disco-like, /google-maps-list-builder — produce the rows this skill fills in
• /list-quality-scorecard — grade the result
• /email-gateway-detection — the mx_class column is a lighter version of what that skill does from DNS
• /cold-email-starter-kit — upload the verified rows