find-first-customers

name	find-first-customers
description	Find and qualify a startup's first customers from current public pain, intent, hiring, migration, and tool-switching signals. Use when Codex is asked to discover early adopters, build a prospect shortlist, research evidence-backed sales leads, score customer fit and timing, or create personalized founder outreach from a startup URL or product description.

Find First Customers

Turn a startup URL or product description into an evidence-backed prospect report. Never invent prospects, quotes, contact details, or source URLs.

Inputs

Collect or infer:

• startup URL or product description
• target geography, if constrained
• B2B, B2C, developer, or local-business motion
• maximum prospect count, default 20
• freshness window, default 90 days

Ask only when a missing constraint would materially change the search. Otherwise state the assumption and proceed.

Workflow

1. Inspect the product and derive an Ideal Customer Profile with pains, triggers, exclusions, likely roles, and search vocabulary.
2. Read references/research-playbook.md. Build queries across at least three relevant public-source families.
3. Search current public discussions. Capture the exact source URL, date, author or organization, a short evidence excerpt, and the signal type.
4. Read references/evidence-schema.md. Store candidates in its JSON shape. Use null for unknown fields; never guess an email or identity.
5. Read references/scoring.md. Score fit, intent, authority, reachability, and freshness independently. Record a one-sentence rationale for every component.
6. Deduplicate candidates by canonical source, public handle, organization, and underlying event. Merge corroborating evidence rather than creating duplicate cards.
7. Resolve this skill folder as SKILL_DIR, then run the deterministic report builder from any working directory:

node "$SKILL_DIR/scripts/build-report.mjs" --input prospects.json --out customer-signal-report

8. Inspect summary.json and report.html. Recheck the top five sources before presenting results.

Quality gates

• Require a working original source URL for every prospect.
• Separate direct buying intent from weak topical interest.
• Prefer recent first-person pain, active migration, budget, hiring, or evaluation signals.
• Mark stale evidence instead of silently treating it as current.
• Use only public professional channels explicitly exposed by the person or organization.
• Keep outreach openers specific to the evidence, low-pressure, and under 70 words.
• Do not automate contact or post/send messages without explicit authorization.

Outputs

The report builder creates:

• report.html — searchable, filterable evidence dashboard
• prospects.json — normalized, deduplicated machine-readable data
• prospects.csv — CRM-friendly export
• outreach.md — personalized opener variants
• summary.json — run metrics and signal/source coverage

Gotchas

• Search-result snippets are discovery aids, not evidence. Open the original page.
• A complaint about a category is not automatically purchase intent. Score the timing separately.
• A company name in a profile does not prove employment. Use the public source's own context and qualify uncertainty.
• Do not infer private emails from names or domains.
• Avoid broad scraping or login-gated sources. Respect platform terms, robots rules, and rate limits.
• Do not include sensitive personal data, protected-class inferences, or non-professional contact details.