prospect-radar

Prospect radar

signal-prospecting researches a market once. This skill keeps watching it. It sets up feeds that surface new companies with a reason to talk, then each run turns the newest items into a short, sourced list and drafts.

Everything lives in one folder (default ./prospect-radar/ in the working folder): feeds.md, prospects/_seen.csv, one prospects/<company>.md per keeper, and drafts/.

Ground rules

1. Every prospect needs a dated, linked reason. A company name in a feed is not a reason.
2. Public business information only, unless the user has connected their own CRM or data tools. Never scrape sites that forbid it.
3. Never guess an email address. Never send anything; drafts are for the user.
4. Run the pre-send checks before any draft (../pre-send-check/references/checks.md, or the pre-send-check skill).

Part 1: Build the feeds (first run, or "tune my feeds")

1. Work out industries, regions and languages, company size, buying roles and the job titles they hire, and competitor or incumbent tools. If gtm-motion.md exists, read it. If the user has named at least what they sell, to whom and where, do not stop to ask: build a first set now, state the assumptions you made for the rest, and ask at most one question at the end to sharpen it. When tuning existing feeds, the feeds themselves show the market: rewrite them now - drop or narrow the broad ones, end with fewer and sharper feeds than before - then ask at most one question.
2. Build feeds using references/prospect-sources.md: one Google News search feed per industry, region and trigger (funding, new digital product, new leader in the buying team, expansion, competitor change), plus job-board searches for the buying roles. Keep each query narrow. Aim for 10 to 20 feeds; when tuning, end with fewer than the set you replaced and state the count before and after.
3. Test every feed with curl -sL -A "Mozilla/5.0" "<url>" (or a fetch tool). Keep it if it returns recent items naming companies in the target market; rewrite or drop it if not. Show a two-line sample from each.
4. Save feeds.md as a table: name, trigger, query in plain words, URL, last tested, hits.
5. Offer to walk the user through Google Alerts for anything personal (a competitor, a key account) and add the RSS links they paste.

Part 2: Each run

1. Harvest. Read every feed's items from the last three days (seven on a weekly schedule). Pull out the company each item is about. Skip companies seen in the last 30 days (prospects/_seen.csv). Aim for 15 to 30 candidates.
2. Qualify. Drop disqualified, out-of-territory and existing-customer companies. If a CRM is connected, skip accounts owned by someone else and note unowned ones; never change ownership. Record every candidate in _seen.csv with the outcome.
3. Enrich the keepers. News and hiring from the last 90 days; the tools their website loads (references/tech-fingerprints.md); the person closest to the problem and one alternate, with sources.
4. Score and pick the top five by fit, freshness of the reason, and whether there is a named person to contact.
5. Check. Run the pre-send checks on each pick. Drop or rewrite as they say.
6. Draft. For each pick: why now (sourced), the angle, who, the checks, and a first touch with one specific observation, one hypothesis and one low-effort ask. Save to drafts/.
7. Report. One line per pick, plus which feeds produced keepers.

Scheduling

To run it every weekday morning, create a scheduled task that runs this skill in the folder (in Claude Code Desktop: Routines, New routine, Local). Weekly works too; widen the harvest window to seven days.

Example

A founder sells a returns-management tool to mid-size online retailers in the UK and Germany. The radar builds 14 feeds: funding and new-leader feeds per country, "launches app" feeds for retail, hiring searches for "ecommerce operations" roles, and a competitor-change search for two incumbent tools. Monday's run reads 212 items, finds 23 companies, keeps 7, and drafts 4 after one is dropped by the checks because the founder emailed them last week.