Terminal
Terminal/Skill
Finds net-new companies from recent trigger events like funding, breaches, or exec hires, with no starting list, each row sourced and ICP-checked.
By Clay
Works with
npx skills add clay-run/clay-skill-creator --skill signal-sourcerThe insight: signals create rows; qualification creates prospects. List-first plays start from accounts whose identity is given — enrichment is the only risk. Here every row is BORN from a noisy recall channel (an event-vocabulary news query), so the row itself must earn existence through three gates the naive build skips: (1) a real, dated event — the channel's date stamps are crawl dates, not event dates, and years-old articles surface inside a one-week window looking fresh; (2) the right entity — a headline names a brand word, not a company; it must resolve to a canonical domain with corroboration before anything downstream spends on it; (3) in-ICP and net-new — an event proves a state change, never fit, and an event at an account already in the user's book is monitoring, not sourcing. The naive build queries news and ships the headline list: rumor-shaped rows.
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 where 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 |
|---|---|---|
| Signal menu | which event types create a prospect for them | ask, and test each one: would the event make someone buy sooner? If not it is trivia |
| ICP guardrails | vertical, geography, size band | non-negotiable. Without them every fired event qualifies and the play degenerates into news clipping |
| Window and cap | how recent counts, and how many qualified rows they want | the past week is defensible, as is 2–3 query variants per signal type. State both |
| The book | existing customers, open pipeline, named accounts | ask — net-new is defined against this list, and without it the play sources their own customers. If they want events on accounts they already know, that is a different skill and say so |
| Owner mapping | territory or segment routing rules | optional. Without it rows deliver unrouted, which is a fine outcome |
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 — a sheet applied silently is a wrong field nobody catches. If there is no sheet, say nothing about sheets — the check is a file lookup, not a question, so run the interview as though the feature did not exist rather than reporting an absence. At delivery, offer to save the answers back (identifiers only — never a token or a password), private and never published — and phrase the offer so it explains itself: "want me to save your answers to a file, so the next person on your team doesn't have to answer these again?"
Run clay whoami; echo "exit_code=$?". If it fails or Clay tools are missing, run
the Clay plugin's setup skill, restart if it says to, and re-run this skill. Tell
the user which workspace you're in. Confirm the detector arm exists before promising
it — read the live catalog, never your memory of it: clay workflows actions list,
grep for the news-query action
(references/detector-mechanics.md has the contract). Note what this surface does
NOT have: no RSS/trigger-source/webhook detector actions and no event filters on
company search — the event-anchored news query is the net-new arm here; the in-app
signal engine is the graduation path, not something this play builds.
Arithmetic before any spend: (signal types × query variants) × ~1 credit per query — quiet queries bill the same; plus ~1 credit per surviving candidate for resolution/qualification enrichment; plus optional premium corroboration (structured funding/jobs lookups, ~6 credits each) for a named shortlist only. State the worst case; wait for approval.
Per signal type, run 2–3 query variants composed as event vocabulary × ICP
qualifier (e.g. funding vocabulary × vertical term; incident vocabulary ×
industry term), each windowed with the tightest relative bucket covering the ask.
Mechanics and the build-once query workflow are in
references/detector-mechanics.md. Discipline per call:
Each result is an ARTICLE; the deliverable unit is a CANDIDATE = (company, claimed
event). Harvest with the source-class rules (full table in
references/detector-mechanics.md):
undated, deliverable only if corroboration
dates it.A harvested name is a brand word with article context, not an identity. Resolve
each candidate with the resolve-company-domain discipline, and start with the FREE
arm: company search with the article's context terms (description keywords ×
location/state) — company search has no name filter, so context terms are how a
collision name is disambiguated, and the returned records carry name, domain, size
band, and location (often enough to kill an off-ICP candidate before any credit is
spent). The paid domain lookup returns confident wrong entities on collision names,
and a plausible knowledge-prior domain can belong to a same-brand entity elsewhere —
BOTH are candidates, not answers: corroborate the chosen domain (enrichment/search
echo: name words + registrable label, never the TLD) before promoting. No corroborated domain → the candidate is delivered in the
exceptions tail as unresolved, never guessed. Generic brand names ("Moss",
"Clay") are exactly where wrong-entity rows are minted — partial-stem matches get
a stated reason or get dropped.
On the RESOLVED entity's own enrichment fields (never the article's claims):
suppressed: existing customer), with a pointer to
monitor-buying-signals for watching them.uncorroborated — source unreachable, and the
delivery says which corroboration mode ran.Failed rows drop with reasons, never silently. Zero qualified rows in a window is a valid, reportable result — an honest zero beats a padded list.
Row: company · domain · signal type · evidence line (≤140 chars, from the source) · event date (+ how derived) · source URL(s) · ICP verdict · net-new check · owner (if mapped). Summary: queries run, articles seen, candidates harvested,
per-gate drop counts, credits spent (measured from run metadata where the surface
exposes it). Then the routing note: this play delivers a POINT-IN-TIME sweep; a
standing version of the same ask should graduate to the in-app signal engine
(native signal subscriptions / signal-triggered workflows) — offer the hand-off
with the arithmetic, never rebuild it as a re-scraping loop.
Ask: "Find healthcare companies hit by a data breach this week — we sell incident
response; 200+ employees, US only. Here are our 60 current accounts." Cost stated:
2 query variants ≈ 2 credits + ~1/candidate qualification ≈ 8 worst case — approved.
Sweep returns 10 articles → harvest: 7 candidates after (entity, event) dedupe
(one breach appeared in 3 outlets — merged), 1 social post dropped, 1 "breach
tracker" roundup harvested for 2 additional names flagged corroboration-required.
Resolution: 6/7 corroborated domains; "Meridian Health" stays unresolved (three
same-name orgs, article context insufficient). Qualification: 1 dropped off-ICP
(38 employees), 1 suppressed (already a customer — noted for
monitor-buying-signals), roundup-harvested names corroborated via primary
notices — 1 confirmed, 1 uncorroborated → dropped with reason. Deliver 3 qualified
net-new rows, each: domain, "breach exposed 310K patient records", event date from
the notification filing, source links, size/geo verdict, owner per territory map.
Drop ledger shows all 4 kills. Offer: make it standing via the native signal
engine instead of weekly re-sweeps.