superdesigndev/treg/recruit

Recruit

A job post in, a ranked shortlist out: people who fit, one line on why, and a way to reach them. Every stage is counted and every contact lookup leaves a vendor trail, written to a run file.

Runs on treg, through its MCP tools or the treg CLI; either is enough. Over MCP, each command below maps onto a tool: treg catalog search is catalog_search, treg catalog get is catalog_get, treg call is call. With neither, treg is not set up: https://treg.to/llms.txt says how. Install only if the human asked you to; otherwise tell them how and stop.

The run

1. Read the job. treg catalog search "linkedin job details", then call the best row with the

post URL. No post URL: use the text the user pasted.

1. Extract criteria from the post, yourself: title variants (the title as posted plus two to

four names the same job goes by), must-have skills, seniority, locations (and whether remote is fine), and the hiring company, which is always excluded. Show the criteria to the user in five lines before spending; adjust if they object.

1. Search with the routed people search, one call per title variant (and per location when there

are several):

`` treg call treg.people.search --data '{"title":"Staff Backend Engineer","location":"Berlin, Germany","country":"DE","keywords":["go","kafka"],"limit":25}' ``

limit is the price dial: most providers bill per row. Start at 25 per query; raise it only when the funnel comes out thin. Read _treg.ignored_filters: a provider that could not apply a filter answered a looser question, so those rows get the hard filters below with no benefit of the doubt.

1. Dedupe by LinkedIn URL (lowercase, no query string, no trailing slash); rows without one by

name + company.

1. Hard filters. Drop, and record a reason for each drop: works at the hiring company, wrong

location, seniority too far off, title from another job family, no profile URL to contact from.

1. Rank, yourself: read each survivor against the criteria and write one line, in the user's

words, on why they fit ("8 yrs Go at a payments scale-up, led a Kafka migration"). No line you can write, no rank. Keep the top N (default 20).

1. Evidence (optional, cheap), for the top E of them (default 10). Gather cited facts, then

re-score: a claim the evidence backs moves a candidate up, one it contradicts moves them down. Cap each candidate's evidence spend (default $0.05) and stop at the cap. Skip a source when the person has nothing for it (no GitHub handle, no posts); never guess a handle.

Read each answer before you cite it: a search result that does not name the person is not evidence. The posts feed mixes in other people's posts; count only items whose author.handle is the profile's own, from the last six months. The job-change check answers a status (NO_NEW_COMPANY means still there). When a source refuses (a provider out of capacity), skip it for the run; its refusal names alternatives, which you may use if they fit under the cap. treg catalog get <id> first if a call is new to you: the catalog changes, and the price and parameters on the row win over this table.

1. Contact, for the top N only:

`` treg call treg.people.email.find.personal --data '{"linkedin_url":"…"}' --header 'X-Treg-Route-Verify: true' --header 'X-Treg-Route-Max-Cost: 0.30' treg call treg.people.phone.find --data '{"linkedin_url":"…"}' --header 'X-Treg-Route-Verify: true' --header 'X-Treg-Route-Max-Cost: 0.30' ``

Both are routed: treg tries providers cheapest per hit first, most misses are free, and the verify header checks a hit in the same call (the mailbox; for a phone, that the line is live). Send only the LinkedIn URL: it is what every provider takes. Never re-send a lookup for the same person; every hit bills, repeats included.

1. Hand back the shortlist and the run file, and state coverage plainly: "214 found, 61 kept,

top 20: 14 emails, 9 live mobiles, $4.12".

The run file

Write recruit-run-<YYYYMMDD-HHMM>.json next to the user's work, and update it after every stage, so a stopped run still says what it did. Money comes from the answers, never from a price list: _treg.charged_micro on each call (micro-USD), _treg.tried for the vendors asked, and _treg.verification.cost_micro for the check.

{
  "job": {"url": "…", "title": "…", "company": "…", "company_domain": "…", "location": "…"},
  "criteria": {"titles": ["…"], "skills": ["…"], "seniority": "senior", "locations": ["…"],
               "remote_ok": false, "exclude_companies": ["…"]},
  "stages": [
    {"stage": "job",     "calls": 1, "charged_usd": 0.0},
    {"stage": "search",  "calls": 6, "rows": 214, "charged_usd": 0.0},
    {"stage": "dedupe",  "in": 214, "out": 163},
    {"stage": "filter",  "in": 163, "out": 61,
     "reject_reasons": {"hiring company": 9, "location": 52, "seniority": 27, "job family": 14},
     "reject_samples": [{"name": "…", "linkedin_url": "…", "reason": "location: Munich"}]},
    {"stage": "rank",    "in": 61, "out": 20},
    {"stage": "evidence", "people": 10, "cap_usd": 0.05, "charged_usd": 0.0,
     "by_source": {"linkedin_posts": {"tried": 10, "found": 7, "charged_usd": 0.0}}},
    {"stage": "contact", "people": 20, "emails": 14, "phones": 9, "charged_usd": 0.0}
  ],
  "candidates": [
    {"rank": 1, "name": "…", "title": "…", "company": "…", "location": "…", "linkedin_url": "…",
     "why_fit": "…", "score_before": 7, "score_after": 8,
     "evidence": [
       {"type": "linkedin_posts", "claim": "…", "source": "anyapi.linkedin.profile_posts_full",
        "url": "…", "charged_usd": 0.0}
     ],
     "evidence_usd": 0.0,
     "email": {"value": "…", "verdict": "valid"},
     "phone": {"value": "…", "verdict": "live"},
     "reach": [
       {"lookup": "email", "endpoint_id": "…", "vendor": "…", "outcome": "miss", "charged_usd": 0.0},
       {"lookup": "email", "endpoint_id": "…", "vendor": "…", "outcome": "hit", "charged_usd": 0.0},
       {"lookup": "email_check", "endpoint_id": "…", "verdict": "valid", "charged_usd": 0.0},
       {"lookup": "phone", "endpoint_id": "…", "vendor": "…", "outcome": "hit", "charged_usd": 0.0},
       {"lookup": "phone_check", "endpoint_id": "…", "verdict": "live", "charged_usd": 0.0}
     ]}
  ],
  "totals": {"charged_usd": 0.0, "emails": 14, "phones": 9}
}

• evidence holds only what a call returned: type (linkedin_posts, github, talks,

papers, job_change, tenure, company_news), a one-line claim, the source endpoint id (search_row for tenure), the url when there is one, and that call's charged_usd. A shared company-news call is charged once, to the first candidate who used it.

• reach is one entry per row of _treg.tried (endpoint_id, provider as vendor, outcome,

charged_micro / 1e6 as charged_usd), then one for _treg.verification when a check ran (served_by, verdict, cost_micro). A skipped attempt keeps its outcome: skipped.

• email / phone are null on a miss. verdict is the check's own word: an email valid,

invalid, catch_all, risky or unknown; a phone live, dead or unknown. A live line is not proof it is this person's number; never call a phone "verified".

• reject_samples: up to five per reason, so the reasons can be checked by eye.
• totals.charged_usd is the sum of every charged_usd above: job, search, evidence and contact.

Before the first list

Say once: contacting people in the EU and UK needs a lawful basis, and anyone who asks to be left alone is dropped for good. Then hand over the list; the user decides how to use it.