Run your first campaign

The cold-start workflow: an owner or founder with no list, no CRM, and no campaign history, taken from "here's my business" to an approved campaign sheet. It chains existing skills in a fixed order with an explicit owner approval between steps. Nothing in this repo sends email - the end state is a drafted 3-email sequence plus a campaign sheet (owner-readable Markdown and a HubSpot-import-shaped CSV) the owner approves on screen.

Chain mechanics follow the router's conventions: each component skill runs by its own SKILL.md, writes into ./runs/ under its own step-prefixed run-id, and hands records.jsonl to the next step by explicit path.

Step 1 - Business context

Check for context/offer.md and context/icp.md. If both exist, summarize them in two sentences and confirm they still describe the business. If either is missing, offer two paths: the closest vertical preset from reference/presets/ (fastest - adopt the archetype, then adjust it together) or the full gtm-context interview. Write the result to context/. Wait for the owner to confirm the context before anything else runs.

Step 2 - Source of record

Where these buyers are already listed depends on the ICP. Read the confirmed context and propose the matching source, by name:

• Local, licensed, or physical-presence businesses (contractors, clinics, dealerships, brokerages, restaurants) live in a public directory - a licensing registry, a professional college, a trade-association "find a member" directory, a marketplace, with 02-apify-maps-discover as the fallback when no structured source exists. These extract via 03 (below).
• Firmographic B2B (software, agencies, funded or hiring companies) has no directory to scrape - the buyer is defined by what the company is. Pull the list instead: 01-prospeo-discover for an ICP-filtered pull, 01-prospeo-lookalike when the owner can name a few good-fit companies, or 04-theirstack-jobs when a hiring signal defines the buyer.

Offer 2-3 candidate sources with a one-line reason each; the owner picks or corrects. Don't open by asking the owner to supply URLs, and don't pull or extract anything before they confirm the source and the cost.

On confirmation, run the matching source skill. For a directory, that is 03-firecrawl-research extract mode with the listing-row schema (03's "Directory and registry extraction" section). Firecrawl bills extraction by tokens (1 credit = 15 tokens), so cost scales with page size: extract the first listing page alone, read the actual charge, and use it as the per-page figure before extracting the rest. For a firmographic pull, that is the discovery skill's own metered call. Either way, state the cost estimate first, follow the source skill's own confirmation thresholds, report the result (N records, credits used), and wait for an acknowledgment before Step 3.

Step 3 - Qualify, free

Run 01-icp-qualify on the extracted records against the confirmed context. This step costs nothing. Report the counts - N qualified, M uncertain, K disqualified - and ask 01's uncertain-gate question: forward the uncertain rows too, or hold them? Wait for the answer.

Step 4 - Rank

Run 05-signal-builder with the vertical-smb calibration on the qualified set: targets ranked, one angle per account, verbatim provenance on every signal. Present the top targets and the segment pattern; wait for the owner to approve the top segment before anything gets written.

Step 5 - Write

Run email-writer for the approved segment: a 3-email sequence built on the top signals and the fallback angle. Drafts only - present them for edits and wait for "drafts approved" before building the sheet.

Step 6 - Campaign sheet and final approval

Run 07-campaign-sheet on the final records.jsonl: campaign-sheet.md (who to contact first and why, signal and approach per row) plus campaign-sheet.csv with HubSpot default import headers. Present both and stop - nothing is send-ready until the owner approves the sheet on screen.

Optional branch, only when the owner wants send-ready addresses: run 06-resolution-email-person on the approved rows (cost-gated - state the estimate first), then regenerate the sheet with the email column filled.

Approval gates (must hold)

• Never send anything. Nothing in this repo sends email. If asked to "just send it", say that plainly and point at the approved sheet and drafts.
• Never spend above the approved estimate. Every paid step re-states its estimate before running; anything above the approved line stops and re-asks.
• Never auto-progress between steps. Every step ends at an owner wait.
• A missing API key stops the affected step: name the layer being skipped and offer the no-key path (a smaller manual pull, or qualify / rank / write on whatever data already exists).

Output

End the run with a one-paragraph recap: the source of record used, N found -> M qualified, the top 3 targets with one-line angles, the sheet path, and elapsed time from context confirmation to sheet.

Reference

• reference/presets/ - vertical context presets (offer.md + icp.md pairs)
• reference/gotchas.md - failure patterns from real cold-start runs
• reference/examples/happy-path.md - one full worked run, from preset to sheet