anysite-company-sourcing
The craft of sourcing companies from Anysite's 70M+ LinkedIn company database (search_sql_companies) - turning a fuzzy ICP into per-field DSL filters that return real matches instead of token soup. Fixes the default failure mode where naive keyword queries return companies from the wrong country, wrong industry and stub pages. Use when the user asks to find/source companies, build a company list, complains that company search results are bad or irrelevant, or needs a target-account universe - "найди компании", "плохие результаты поиска компаний", "source accounts". For people at those companies use anysite-people-sourcing; for funding-stage filters use crunchbase; for pushing into a CRM use anysite-crm-prospect.
anysiteio · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add anysiteio/agent-skills --skill anysite-company-sourcingAbout this skill
name: anysite-company-sourcing description: The craft of sourcing companies from Anysite's 70M+ LinkedIn company database (search_sql_companies) - turning a fuzzy ICP into per-field DSL filters that return real matches instead of token soup. Fixes the default failure mode where naive keyword queries return companies from the wrong country, wrong industry and stub pages. Use when the user asks to find/source companies, build a company list, complains that company search results are bad or irrelevant, or needs a target-account universe - "найди компании", "плохие результаты поиска компаний", "source accounts". For people at those companies use anysite-people-sourcing; for funding-stage filters use crunchbase; for pushing into a CRM use anysite-crm-prospect.
Company Sourcing
linkedin/search/search_sql_companies searches 70M+ companies and is the single
best bulk company tool in the catalog — and its naive use is the single most
common source of garbage lists. Measured live, same intent, same day:
{keywords: "AI startup San Francisco"}→ 1 relevant of 5: a Tel-Aviv gaming-data firm, a Sydney fintech and a Ho-Chi-Minh beauty e-commerce all matched — each had the token "startup" somewhere in its description and "San Francisco" among secondary office locations.- The same intent as structured filters (below) → 5 of 5 genuine SF AI companies in the right size band.
The difference is the whole skill. Never ship results from a naive query.
Why naive queries fail (mechanics, not opinion)
keywordsmatches whole words across ALL text fields — name, description, specialities, hashtags and the locations array — so "San Francisco" as a keyword still matches a Hanoi company that lists an SF sales office inlocations. (Whole-word matching removed the inner-substring noise, but not the wrong-FIELD problem — HQ is a separate field, use it.)sortchanges ordering, not the candidate set — a naive keyword query is still not an ICP list. With relevance (the default), "AI startup San Francisco" surfaces Startup Weekend AI, Bitcoin AI Startup Lab, 1-employee shops and a Phoenix-HQ company (all measured) — they score high because the words sit in their NAME. The fix is per-field decomposition below, not a sort flag.- Millions of company pages are stubs. Without hygiene filters they dominate.
employee_count_rangecan contradictemployee_countin the same record (measured: 305 employees with range "11-50"). Never filter or segment by the range string.
The method: decompose intent into fields
Take the user's ICP sentence apart and map each fragment to its OWN field. No ICP in the
request → take it from the ICP section of anysite-gtm-profile when it exists (and say so
in one line), including its exclusions.
| Intent fragment | Field | Notes |
|---|---|---|
| "based in X" | country_hq: ["US"] + headquarter_location: "\"san francisco\"" | token-aware; NEVER the locations field — that matches branch offices |
| "present in X" (offices count too) | country_any | this is the only right use of the locations array |
| "does AI / fintech / logistics" | specialities OR industry_name OR description together, not specialities alone | industry array wants URNs; industry_name resolves labels — see the specialities caveat below |
| "in the orbit of company X" | similar_organizations: "\"fsd_company:<id>\"" | queryable filter, not just an output field — the reverse-graph expander; see below |
| "startup / SMB / enterprise" | employee_count_min / employee_count_max | integers; ignore employee_count_range entirely |
| "founded recently" | founded_on_min | year |
| named company lookup | name or alias DSL + exact verification | never trust first hit; a known domain → companies/resolve (anysite-mcp) |
| always, every query | is_active: true, has_website: true, min_description_length: 100 | the hygiene trio kills stubs and dead pages |
| ranking | sort = relevance (default) or last_modified | relevance for sourcing; last_modified for "what's new since last run" (monitoring). Scoring weights a term by field: name 5× > specialities/hashtags 3× > short_description 2× > long description 1×, length-normalized; ties broken by recency. Caveat: the score is built only from keywords/name/specialities/description — a query filtered ONLY by non-text fields (e.g. just industry + employee_count_min) has nothing to score, so it falls back to recency order |
Industry labels are often wrong or empty (auto-created LinkedIn pages especially). When
missing a company matters more than extra noise, run a second query with the description
terms and NO industry filter, merge the two (merge_data with dedupe_by: ["urn"]) and
decide each row with one yes/no question — the wide-net method in anysite-crm-lookalikes.
DSL in every text field: whitespace = AND, | = OR (no spaces around it),
"phrase" = exact phrase / substring, -token = NOT. Example:
specialities: "\"artificial intelligence\"|\"machine learning\" -agency".
IRON RULE: quote every multi-word alternative in an OR chain. Whitespace
binds tighter than | — name: "Level Infinite|Proxima Beta" parses as
(Level) AND (Infinite|Proxima) AND (Beta) and returns 0 (verified live on the
sibling people endpoint, same parser). Lint before sending: a space inside an
OR alternative without quotes → fix first.
Matching semantics (changed — this is now the biggest lever):
- A bare word matches as a WHOLE WORD, not a substring (verified live:
keywords: "sdr"returns "SDR Academy", "SDR Foundry", "Vida SDR" — companies with the standalone word, and no longer "adviseur"-style inner hits). Case doesn't matter. - No stemming.
integrationdoes NOT findintegrations— write both:integration|integrations. Same for singular/plural and verb forms. - Quotes = substring (the old behaviour, kept). `