Prospecting System

Turn your 5 best customers into your next 500 prospects. This skill does no prospecting itself: it runs the 20 stage skills in order, moves files between them, tracks spend, and stops at the gates where a human decides.

Intake (ask all of it in one message)

1. Best customers — 5 to 10 domains, plus anything known about each deal (size, champion title, how they found you). Saved as 00-customers.csv.
2. Your website — for the ICP stage.
3. Target — how many send-ready leads (default 500). It is a ceiling, not a quota: if 180 qualify, deliver 180 and say why.
4. Spend cap for this run, in USD — required. No default. Covers every paid call: follower scrapes, DiscoLike, Prospeo, enrichment credits, email verification.
5. Lanes — accounts (lookalikes + search), audiences (followers, engagers, communities, events), or both (default).
6. Where copy goes — Smartlead, HeyReach, another sender, or CSV only. This system never sends.

Then create runs/<customer-slug>-<YYYY-MM-DD>/ with 00-customers.csv and state.json.

The stages

Read the stage's SKILL.md before running it. Its "Pipeline contract" block is the handoff; the rest of the file is how to do the work.

Stages 3–6 (audience lane) and 3+7 (account lane) are independent until stage 8 merges them. A lane the user declined is marked skipped.

Operating rules

1. Spend cap is hard. Before any paid call, estimate its cost, add it to spent_usd, and stop if the total would pass spend_cap_usd. Ask the human; never work around it, never split a purchase to slip under it.
2. Dedupe before you pay. Against the run so far, excluded_domains, and the do-not-contact list.
3. Cheap before deep. Free counts and the ICP judge run before enrichment; enrichment before research; research only on leads that will be contacted.
4. Unknown is not negative. A blank field stays blank and is scored as unknown. Never turn missing data into a zero or a guess.
5. Fit, intent, and priority stay in separate columns, each with its evidence.
6. Every signal has a source URL and a real event date, or it is not written.
7. Never pad. Fewer qualified leads than the target is a result; report the constraint.
8. Drafts only. Nothing in this system uploads to a sender or sends a message.
9. Bulk per-row work goes to Haiku sub-agents (judging, classifying, personalizing), in batches, after a human has approved a sample. Strategy, ICP, and copy frameworks stay on the main model.
10. Human decisions are proposed, not edited: the ICP yaml, the judge prompt, score weights, segments, copy frameworks.

state.json

{
  "run": "acme-2026-01-15",
  "target_leads": 500,
  "spend_cap_usd": 150,
  "spent_usd": 0,
  "stages": {
    "01-customer-patterns": {"status": "done", "rows_in": 5, "rows_out": 5, "cost_usd": 0, "finished_at": "2026-01-15", "note": ""}
  }
}

Stage keys are <NN>-<skill-name> (04-audience-discovery, 08-contact-enrichment, ...). status is pending, done, skipped, or blocked (with the reason in note). Update it when a stage finishes, then run:

uv run scripts/check_run.py runs/<run>

It validates files and columns against this contract and prints the stage funnel. Fix what it reports before starting the next stage.

Resuming: read state.json, run the checker, continue from the first stage that is not done or skipped. Never redo a paid stage whose output file exists.

Lead columns

From 06-audience-qualified.csv and 08-leads.csv onward, every lead file carries these columns; stages only ADD columns.

Account files (03, 07) need domain; signals need the columns in /buying-signals.

What "done" looks like

After stage 16: a ranked, segmented, verified lead list with first-touch email, LinkedIn note, and follow-ups per lead, plus a run summary — leads by source and tier, spend against cap, and what limited the count. After stage 20: a dated review with proposed targeting changes for the next run.