Terminal
Terminal/Skill
Builds a verified-email lead list from your ideal customer profile, saves it as an audience segment and drafts an unsent email sequence.
By Clay
Works with
npx skills add clay-run/clay-skill-creator --skill icp-outbound-campaignThe insight: the strict ICP on the source run found 141 people for a target of 200, so this skill widens in declared tiers and fills them in order rather than loosening the definition. The source run targeted CISOs and heads of security at fast-growing software companies in the US and Canada on 2026-09-28:
| Tier | Definition | People found |
|---|---|---|
| 1 | persona titles, core industry, 51–5,000 employees, 20%+ 12-month headcount growth | 141 |
| 2 | same titles and growth, adjacent industries | 52 more (193) |
| 3 | same titles, industries and growth, 11–50 and 5,001–10,000 employees | 99 more (292) |
The strict definition alone could not reach 200 — and it lost more on the way: 12 titles that matched the search but do not buy (vendor "Field" titles, deputies, program managers), 2 whose company resolved to a junk domain, 24 with no verified email, and 24 who were a second leader at a company already on the list. The run kept 200 of 230 eligible, filled tier 1 first (116), then tier 2 (38), then tier 3 (46). How big any other ICP's pool is was not measured; that is why the skill counts tier 1 before promising a number.
Four more things follow, and the first three bit the run that produced this skill.
The ICP is the installer's, and the workspace may hold the wrong one. The source workspace's saved business context described a different company than the seller. An ICP read from context and used silently would have targeted someone else's buyers. So this skill asks for the ICP every time, and shows any saved ICP only as a starting point to confirm.
Title similarity matching over-reaches in a predictable direction. A title search expands into related titles, and some of them never own a budget — on the source run, asking for "CISO" also returned Field CISOs at vendors, deputies, fractional CISOs and advisors. Every persona has its own version, so the skill shows the titles that came back and asks which to drop before anything paid runs.
A Sequencer follow-up is not a reply. The sequence stops for a lead the moment they reply, so the second step only ever reaches people who did not answer. Anything meant for people who said yes is a message a human sends from the thread, not a sequence step.
A personalization data point is only worth paying for if it reaches the email. The source run personalized on nothing beyond the fact the search filtered on, so it needed no extra enrichment. An extra data point — a recent post, a funding round, a news story — costs a lookup per lead, returns nothing for some of them, and can be wrong about the company. So the skill asks whether the installer wants one, picks the enrichment by testing it on real rows, and writes copy that still reads naturally for the leads where it came back empty.
Nothing here ships with a value. Each one is the installer's, not the author's: ask for it, never substitute a plausible default, and if an answer does not exist say which step becomes unavailable rather than guessing. Where a default IS defensible it is named below, and using it means saying so in the output.
| Input | What the installer supplies | If it is missing |
|---|---|---|
| Buyer persona | the job titles that count as the buyer, and the seniority levels | no default — Step 3 has nothing to search for |
| Company profile | industries, employee size range, and any company signal that defines fit (headcount growth, hiring for a role, technology used, headquarters location) | no default — ask. A search with titles and no company profile is too broad to send to |
| Geography | where the people must be based | no default — ask |
| Target count | how many leads they want in the finished segment | the author recommends 250–300 for a first test campaign; recommend it, let the installer decide, and never pick a number for them |
| Exclusions | companies or segments never to contact (customers, competitors, an existing Audiences segment) | everyone already in the workspace's Audiences is excluded regardless; ask for anything beyond that |
| Title exclusions | titles that match the search but never buy | built from the search results in Step 4 and confirmed by the installer; never assumed |
| Widening order | which ICP criterion to relax first if tier 1 falls short | asked in Step 3 only if tier 1 falls short; never widen silently |
| Leads per company | how many people one company may contribute | the source run used 1, a run-time default rather than a stated rule; say it is borrowed |
| Personalization data points | any extra fact per lead or per company they want the emails to use, and how recent it must be | none is a complete answer — the campaign personalizes only on name, company and the ICP fact the search filtered on |
| Audiences fields for data points | the contact fields that will hold each extra data point | read the Audiences field list and propose a match; if none fits, ask whether to create one (a write, approved in Step 6) |
| Seller and offer | who is sending, what is offered, and what the recipient has to give to accept it (a call, a form, access) | no default — Step 9 cannot write copy |
| Product facts and proof | one or two sentences on what the product does, and any customer result the installer is cleared to cite | copy runs with no proof line. Never invent a customer or a number |
| Lead-source tag | the value written to the lead-source field so this batch forms its own segment | no default — without it Step 8 cannot build a segment that holds only this run |
| Lead-source field | the Audiences contact field that holds the tag | read the Audiences field list, show the match, ask them to confirm; if none exists, stop and ask them to create one in the Clay app |
| Save-to-Audiences workflow | an existing workflow that takes a CSV and upserts contacts, or permission to build one | Step 8 builds one after the Step 6 approval |
| Follow-up delay | days between the first email and the follow-up | the author recommends 3–4 days; if they have no view use 3 and say it is the author's |
| Third email | whether to add a third email, and its delay | two emails is the default. The author's experience is that anything after email 2 has a very low chance of a response; if they want a third, it goes 7 or more days after email 2, and say so |
If an answer sheet is present beside this skill, load it and ask only for what it does not cover. A partial sheet is normal; a value it is missing gets asked for on its own rather than restarting the interview. Say which values came from the sheet before using them. If there is no sheet, say nothing about sheets.
Run clay whoami. It must return a user and a workspace; name the workspace out loud. If it fails,
say which component is wrong and the one command that fixes it (clay login), and stop. Do not
install, upgrade or fetch anything to repair the platform.
Then say, before anything else: "I'll ask you about your ICP first, then search Clay for people who match, find their work emails, add any extra details you want to personalize on, save the kept leads into your Audiences as a new segment, and create a draft campaign. Nothing is sent or launched — you do that in the Campaigns UI."
Do not start a step before the steps above it have their answers. If a declared input is missing, ask for it — never assume a default and continue.
The ICP is the installer's judgment about their own market. Ask for it; never infer it from the workspace, the seller's website, or an earlier campaign. One question per message, in this order, each with a one-line example so the installer knows the level of detail wanted:
Before question 1, run clay campaigns options business-context. If it holds an ideal customer
profile or buyer personas, show it and ask whether it applies to this campaign — it may describe
a different company or an old ICP. Use it only on a yes, and only as the starting point for the
questions above.
If an answer is vague ("fast-growing", "mid-market", "tech"), ask once for the number or list behind it. If they have no view, offer the closest searchable default from the live query reference, name it, and record it as borrowed — for example, the reference maps vague growth wording to more than 20% headcount growth over 12 months.
Then play the ICP back in one short paragraph — the titles, seniority, company profile, geography, exclusions and target — and ask "is that your ICP?". Stop until they say yes. This is the Step 1 halt.
clay audiences list --entity-type people, following cursor to the end. Name any segment that
looks like an earlier run for the same ICP.clay campaigns list --search with the seller's name. Name any empty draft that looks like an
earlier attempt.Fetch the live query reference first and write it to a file; never restate it from memory:
clay searches query-mode reference | jq -r '.reference' > ./clay-search-reference.md
Read its grammar, operators, people guardrails, the people and experience field catalogues, and any topic section the ICP needs (location, company size, dates and growth, products and services). Map each ICP criterion to a field. Every enum value must come from the reference's own list — an industry or size band that is not listed fails validation.
The shape, with every value from the confirmed ICP:
select from people
where experiences.any(is_current = true
and job_title is_similar_to (<persona titles>)
and seniority in (<seniority levels>)
and company.industry in (<industries>)
and company.company_size in (<size bands>)
and <company signal, e.g. company.employee_growth_12mo > 1.2>)
and location_country in (<geography>)
and clay.exclude_people_identifiers(@audience_segment("ALL"))
Keep every company condition inside one experiences.any(...). ALL excludes every contact already
in the workspace's Audiences; add one clay.exclude_people_identifiers(@audience_segment("<id>"))
per extra exclusion segment the installer named, resolved from clay audiences list.
If an ICP criterion has no search field, say so plainly — name it, say what the search will use instead (the closest field, or nothing), and offer to check it per lead as a data point in Step 5. Never drop it silently.
Before creating anything, read the search back in plain words — "people titled VP or Head of
Sales, Director and above, at software companies with 201–1,000 employees that grew headcount more
than 20% in the last year, based in the UK, not already in your Audiences" — and ask them to
confirm. Then create with clay searches query-mode create --query and page with
clay searches query-mode run <searchId> --limit 100 while hasMore is true. Keep each tier's rows
in its own file — never merge with a shell glob that can pick up unrelated JSON files; that
happened on the source run and polluted the merge.
After each run that returns periodQuota, check that the planned pull leaves at least 15% of
limit; if not, stop and ask.
Show the first page as a sample — the 10 most common titles and a handful of companies — and ask whether these look like their buyers. This is the Step 3 halt, and it is where a wrong industry or an over-broad title shows up before anything is spent.
Widen only if tier 1 cannot reach the target. Aim for roughly 1.4× the target in candidates,
because the source run lost about 30% between search and final list. If tier 1 falls short, say how
many it found and ask which criterion to relax first — an adjacent industry, a wider size band, a
looser signal, a neighbouring geography. Each widening is its own tier and its own search; label
every row with its tier. Dedupe across tiers on clay_profile_id.
Rows carry clay_profile_id, name, first_name, last_name, linkedin_url, location.name and
matched_experiences[0] with title and company. On the source run query mode returned
linkedin_url; if a row lacks it, the email step still runs on name, domain and company.
matched_experiences[0].title is empty or matches the confirmed exclusions
(case-insensitive, whole-phrase).Say the counts: found, removed by title, remaining, unique companies.
Ask one question, with examples, and accept "none" as a complete answer:
"Beyond their name, company and the ICP fact we searched on, is there anything else you'd like the emails to reference? Clay can find things like a person's recent LinkedIn posts, a company's recent news, or its latest funding round — or something else specific to your pitch. Name as many as you like, or say none."
If they say none, skip to Step 6. For each data point they name, do the following. Do not name a provider or function to the installer until you have checked it exists in their workspace.
clay routines list and follow cursor to the
end. Match on each routine's input schema (clay routines get <id>), not its name — a routine
whose inputs the kept rows can supply (profile URL, full name, company domain, company name).
Prefer source: managed, then custom routines. Only if none fits, search the action catalogue with
clay workflows actions --help to find the listing and schema commands, and read the action's real
input schema. Never carry a function name or price in from memory; read them from this workspace.estimatedCreditCost from clay routines get (the list call omits it), or the
action's own cost and paymentType. Read parameter descriptions for per-result pricing: a step
that bills per item returned cannot be totalled before it runs, so give its per-item price and set
a cap on items per lead.clay audiences fields --help for the
commands) and propose an existing contact field for each data point, or a new one to create. For a
long raw result (a post body, an article), keep a short extract — one or two sentences plus the
date and source link — so the field stays readable and the copy step is grounded.Then play back each data point in one line: what it is, which enrichment, per-lead or per-company, test hit rate, cost, and the field it lands in. This is the Step 5 halt; wait for a yes before the gate.
Read declared costs by id — the list call omits them:
clay routines list # find the managed company-domain and work-email functions by input schema
clay routines get <routine id> # estimatedCreditCost.perRun
On the source run the managed Company Domain function declared 1 credit per run and Work Email declared 3 per run. Read them again; the installer's plan may differ.
Run the domain and email lookups on 10 rows first (the inline form, --input with up to 100
items), and show the 10 results: company, domain found, email found or not. The Step 5 tests already
cover any personalization enrichment.
Then one message, and stop:
vercel.app, netlify, github.io, herokuapp, wixsite, notion.site, or a social-network domain);This single gate is both the spend approval and the write approval.
Company Name. The bulk result is a JSONL file at resultUrl from
clay routines runs get <runId> --bulk; each line is {"id", "status", "result": {"Domain": ...}}.
Drop the junk domains listed in Step 6 — the person is skipped, not guessed.Full Name, Company Domain,
Company Name, and Social Profile URL from linkedin_url when present. Read
result["Work Email"]. An empty result: {} means no verified email — skip the person, never guess
a pattern.clay routines runs get <runId> --bulk --wait 600. Report found and not found for each.From people with an email, walk the ranked list and keep up to the declared leads-per-company, deduping on email, until the target count.
Write a CSV with full_name, first_name, last_name, title, company, location, work_email, linkedin_url, lead_source, plus one column per approved data point.
Create any approved new Audiences fields for the data points, confirming the command on the
installed CLI with clay audiences fields --help.
Find the save-to-Audiences workflow — clay workflows list, then clay workflows graph get <id>
for a CSV-upload trigger feeding an upsert-audiences-record tool node on CONTACT. If one exists,
read the node with clay workflows nodes get and confirm its trigger fields match the CSV headers,
including the data-point columns. If it does not map them, say so; updating an existing workflow's
mapping is a write the installer approves here, never a silent edit.
If none exists, build one (approved in Step 6). Confirm the command forms on the installed CLI
first — clay workflows --help, clay workflows nodes --help, clay workflows triggers --help,
clay workflows actions --help — never trust a form written here. Two nodes, one edge:
csv_upload trigger whose fields are the CSV headers;upsert-audiences-record: entity type CONTACT; lookup on email and
profile URL; record fields name, first name, last name, title, job title, company, email, profile
URL, location, the installer's lead-source field, and each data-point field; remove-null-values
on, so a blank never overwrites a populated field.Tool-node outputs are read at $.result, and a tool node does not echo its inputs. Say out loud:
"I'm writing a line into the workflow's description so this can be traced back to the listing."
Apply the marker only when this installed copy carries marketplace_slug and
marketplace_revision; the line is Sourced from marketplace skill: <slug>@<revision>. Read the
description first: null → write a description then the marker; a description with no marker →
append; this same marker → write nothing; any other marker → stop and report a conflict. If the
identity fields are absent, skip the marker and say attribution is unavailable. Read it back.
Upload and run: clay workflows triggers csv upload <triggerId> --file <csv>, then
clay workflows triggers csv run <triggerId>.
Create the segment with clay audiences create --entity-type people, filtering on the lead-source
field equal to the tag AND email not empty. Copy the filter's key and dataPath from an existing
segment's clay audiences get output rather than guessing them. Name the segment after the ICP and
the count.
Verify: clay audiences records search-count --entity-type people --audience-id <id> equals the
CSV row count. If it does not, say how many landed and stop.
clay campaigns create --name "<seller> — <offer> — <ICP label> (<date>)".clay campaigns update <id> --input '{"leadBaseSegmentId":"<segment id>"}'.clay campaigns audience <id> — use only fields by their exact ids. A data-point field will not be
reliablyPopulated when some leads lack it; that is expected and is why it gets a fallback.clay campaigns context resolve <id>. If the attached business context describes a different
company than the seller, say so and do not let AI snippets draw on it.clay campaigns sequence edit (read --help for the op schema):
{{ai:|Opener|Write one sentence referencing this recent item: {{lead:<fieldId>|<label>|fallback=none}}. If it is none, write a one-sentence opener from the ICP fact instead, without implying research.}}
The snippet may only restate what the field says; it must not infer opinions, results or intent
the field does not contain.clay campaigns sequence spam-check, then clay campaigns variants preview-sequence on one lead
that has each data point and one that does not. If a preview lead's company does not look like the
ICP, say so.No sender accounts are set, nothing is launched, and no test email is sent.
Report the confirmed ICP in one line, the segment name and count, each data point with its hit rate, the campaign name, the final copy with readable placeholders, the coverage line, any ICP criterion the search could not express, and three things the installer does next: add sender accounts and launch in the Campaigns UI; confirm any proof line is cleared to cite; and reply by hand to anyone who says yes, because the sequence stops at their reply.
At delivery, offer: "Want me to save your ICP and settings to a file alongside this? It isn't part of the skill — it's a note of what you told me: your titles, company profile, geography, exclusions, data points and offer. Next time you won't re-answer these, and a teammate who has it can run this without knowing your setup. It stays with you, is never submitted or published, and holds no passwords or keys."
People titled VP or Head of Sales, Director and above, at software companies with 201–1,000 employees that grew headcount more than 20% in the last year, based in the UK and Ireland, excluding current customers and everyone already in Audiences. Target: 150 leads, one per company.
| Data point | About | Window | Enrichment chosen | Test hit rate | Final hit rate | Field |
|---|---|---|---|---|---|---|
| Latest funding round | company | last 12 months | managed company funding routine, once per company | 4 of 10 | 61 of 150 | Latest funding |
| Recent LinkedIn post | person | last 30 days | profile-posts action, capped at 3 posts per lead | 3 of 10 — below half | dropped by the installer | — |
| Group | Leads | Share | Definition |
|---|---|---|---|
| Tier 1 — confirmed ICP | 98 | 65% | the ICP exactly as confirmed |
| Tier 2 — size band widened | 52 | 35% | same, 1,001–5,000 employees, widened on the installer's say-so |
| Not kept — no verified email | 19 | — | skipped, never guessed |
Titles kept: 61 VP of Sales · 47 Head of Sales · 42 Director of Sales.
Subject: Pipeline review for Northwind
Hi Dana,
Congrats on Northwind's Series B in March — a round like that usually means the sales team grows faster than the pipeline data behind it.
I'm with Contoso. I'd like to offer a pipeline review for Northwind, on us: a 30-minute call where we look at your last two quarters of opportunities and show where deals stalled and why.
There's nothing to buy, and you keep the findings either way.
Want me to set one up?
Sam
212 found across two tiers · 9 removed by title · 169 with a verified email · 150 kept after one per company, filled tier 1 first · funding found for 61 of 150 · 0 already in Audiences · draft campaign, not launched.
A good run starts with an ICP the installer confirmed in their own words and a search they approved in plain words, and ends with a segment whose count equals the CSV exactly, a coverage line that accounts for every person found — removed by title, no email, over the per-company limit, or kept — and a draft campaign whose preview reads naturally for a real lead who looks like the ICP. Tier 1 fills first, and any widening was asked for and appears as its own row. Every personalization data point shows the enrichment chosen, why, its test and final hit rates, and a preview for a lead without it that still reads well. A thin run says so: "your ICP found 60; you asked for 200; I stopped before widening", or "posts came back for 3 of 10, so I dropped them." A bad run looks finished and hides one of these — an ICP the agent inferred, a criterion silently dropped from the search, a segment count that does not match, a tier widened without asking, a non-buyer title in the list, a guessed email, an opener about the wrong company, a data point with no fallback that drops leads from enrollment, or a follow-up written as if the person had replied.
A GTM ops lead at a compliance-automation vendor asked for 200 leads in their ICP: CISOs and heads of security at rapidly growing software companies, then a two-email campaign offering a free security assessment. The saved business context in the workspace described a different company, so it was not used. Asked where the leads should be based, they chose the US and Canada; "rapidly growing" had no number behind it, so the reference's default of more than 20% 12-month headcount growth was applied and recorded as borrowed. The workspace already held two earlier lists and two empty drafts from the same vendor; they chose a fresh build excluding everyone already in Audiences. Tier 1 returned 141, so tiers 2 and 3 were added (292 total). Title cleanup removed 12 — Field CISOs, deputies, program managers — leaving 280 across 252 companies. No extra personalization data point was used on this run; the copy personalized on first name, company and the growth the search filtered on. The domain lookup returned 250 usable domains (two junk hosting domains dropped); the email lookup found 254 of 278. One per company left 230, and the first 200 in tier order were saved to Audiences through a CSV-to-upsert workflow and gathered into a new segment of exactly 200. A draft campaign was created on that segment with a three-day follow-up. The first spam check scored 56 ("fair": no spintax, "no cost" flagged); after adding spintax and changing "no cost" to "on us" it scored 100. The preview lead worked at a security vendor, which was flagged. Nothing was launched.