Cold email pipeline

Six stages over a directives folder the user configures once. The stages that touch data call Jell capabilities; the stages that judge (grading, writing) run on you. Two gates: the final list and copy, and the first live send. Everything between them runs on its own.

Research → List → Grade → Verify → Write → Send
  brief    raw    keep    clean   per-    inbox the
           leads  ICP     list    lead    workspace
                  only            copy    owns

Raw lead lists are 30 to 50% wrong-ICP. Grade and Verify exist so the campaign does not email agencies, dead sites and wrong-geo firms, tank the reply rate and burn the domain.

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.
• Check directives/ exists in the working project. If not, copy the directives/ folder next to this file into the project and ask the user to fill icp.md, lead-source.md and email-copy.md. Those three are mandatory; the others have sensible defaults. When no human is in the loop (a scheduled or delegated run), fill them from the brief you were given, say so in the files, and treat the caller's budget as the approval at both gates.
• Run the free email.inboxes. The campaign sends from an inbox the workspace owns on a verified domain. If there is none, stop and run the cold email infrastructure setup first (domain.search, domain.register, email.domain, domain.dns, email.inbox: the guide at https://jell.run/use-cases/cold-email-infrastructure), each purchase behind its own confirmation.
• Inspect people.search, company.search, contact.find, contact.verify, web.scrape and email.send once and show the prices (inspect lists every input field with its allowed values). Quote the campaign as a whole before stage 2. Over the CLI there is no batch_run: loop run with max_cost on each; batch_run with max_total_cost is the MCP form.

Stages

1. Research (once per ICP)

Write directives/research-brief.md: what these buyers care about, the daily friction, the words they use, the events that make now the right time. Use your own web search and fetch when you have them; web.search and web.scrape when you do not or a site blocks you. Every claim in the brief carries a source URL. Reused across campaigns.

2. List

Translate the user's sentence through directives/lead-source.md into people.search filters:

jell run -c people.search -i '{"titles":["founder","managing director"],"industries":["financial services"],"locations":["Lyon, France"],"company_headcount":["1-10","11-50"],"limit":25}' --max-cost 0.35 --wait 120 -o out/leads-raw.json

company_headcount bands are 1-10, 11-50, 51-200, 201-500, 501-1000, 1001-5000, 5001-10000, 10001+ and seniority is entry, senior, manager, director, vp, cxo, partner, owner; anything else is refused before routing. The quote is $0.175 a 25-result page plus $0.007 a profile in full mode ($0.35 at 25, $1.40 at 100), and the quote is the floor. Rows carry first_name, last_name, name, headline, current_title, current_company, linkedin_url, location; email is always empty and there is no web domain, which contact.find needs. Or start from accounts (the hiring-signal-outbound skill, or company.search with the query and the posting's location) and search people inside them: one people.search with up to six LinkedIn company URLs in companies costs one page ($0.175 plus profiles) instead of one page per company. Get the domain and the headcount from company.search per account (domain, employee_count; the range field is the company's declared bracket and often disagrees, filter on the count), and drop rows with an empty current_title or current_company or a truncated last name. Apply negative-keywords.md at search time where a filter supports it. Output: out/leads-raw.csv, after that cleaning.

3. Grade

First drop the rows whose title or posting is not the role (a job search returns Marketo developers and consultancies alongside the growth engineers), then, for each remaining company, read its website and assign a category from directives/qualification.md: ICP to keep; agency, saas, freelancer, enterprise, wrong-geo, dead-site, unclear to drop or hold. Add categories to the rubric when the ICP is itself software (a consumer or below-band row is not unclear). Fetch with your own tool when you have one; web.scrape ({"url": ...}) when you do not. Pre-classify with the negative keywords, then read the fetched text yourself for every unclear row and a sample of ICP rows. The rubric judges, not the keywords. Output: out/leads-graded.csv, all rows kept with category and category_reason, so the user can audit the filter.

4. Verify

On ICP rows only:

1. Free dedupe first: jell history --file people.txt on the LinkedIn URLs (there are no emails yet) says who this workspace already contacted (contacted) and which paid results exist (reusable, with the run id: reuse them instead of paying again). seen alone means a run touched the URL and is not a reason to drop.
2. Email missing: contact.find with first name, last name and company_domain. No-match is released, not billed; the result is absent from the receipt, not null. Run history --file emails.txt once more on the found addresses before the send.
3. Every email: contact.verify. Keep valid; drop invalid; hold catch_all and unknown in their own column.
4. One row per company unless the user targets named individuals; two contacts per company at most.
5. Clean merge fields: strip legal suffixes (Inc, Ltd, SARL, Lda), fix ALL-CAPS names. A broken merge field is an instant blast tell.

Output: out/leads-verified.csv.

5. Write

Per-lead copy from directives/email-copy.md fused with the research brief. Rules: the first line is written from that lead (role, company, one site detail); the angle follows the tier in the directive; one call to action; plain text; 120 to 160 words before the signature and the opt-out line; the product facts match the live site (verify, never invent a feature or a trial term). Write the J+0, J+4 and J+10 rows per lead with a stage column.

Output: out/campaign.csv with to, first_name, company, tier, stage, subject, body.

Gate 1: show the count, the category breakdown, the verification breakdown, three sample emails, and the quote for the send. Wait.

6. Send

1. Send one test to the user's own address from the campaign inbox and ask them to check it lands in the inbox, not in promotions or spam.
2. Gate 2: the user confirms the first live batch.
3. Send from the owned inbox, plain text, with a reply-to and an unsubscribe line:

jell run -c email.send -i '{"inbox_id":"<id>","to":["[email protected]"],"subject":"...","text":"...","reply_to":"[email protected]","unsubscribe_url":"https://yourdomain.com/unsubscribe"}' --max-cost 0.01

inbox_id is the inbox address as email.inboxes lists it; to takes a list (a single string is accepted). unsubscribe_url is a page the user owns; when there is none, leave it out and rely on the opt-out sentence in the body. A send cost $0.0035 when this was written; send one to yourself first, then read it back with email.messages on the same inbox to prove both directions work. Ramp on a fresh domain: 5 to 10 a day for the first three days, 15 through day seven, then 25 to 30. Never burst: spread a day's sends across the day. Log every send to out/sent-log.json as {"run_id", "message_id", "to", "subject", "stage", "sent_at"}; the sdr-daily skill matches replies on to and subject and threads follow-ups on message_id.

Rules

• Inspect once per capability, quote the whole campaign at stage 2 and the send at Gate 1. Stop before a batch over about $1 unless the user asked for that volume.
• max_cost on every run, max_total_cost on every batch_run (MCP). Per-result capabilities bill the requested limit even when fewer rows come back.
• Only email people the user is entitled to contact under the law that applies to them and to the recipient. B2B first-party outreach with an identifiable sender and a working opt-out is the floor; the user decides what their jurisdiction needs and the skill does not lower it.
• Never send from a shared or unverified domain. Never send HTML, images or tracking pixels in a first touch.
• Never fabricate a product fact, a number or a first line. If the research does not support a hook, use a plain one.

Output

The four CSVs in out/, the sent log, and a summary: leads raw, kept after grade, kept after verify, sent, total charged as the sum of every run's billing.charged (not a wallet difference: the balance is shared by every session on the key), with run IDs. Then hand the loop to sdr-daily.