Prospect Builder

Hard rules — read first

1. Landbase API is the source of truth. All enrichment, dedupe, and contact/company lookups go through Landbase first.
2. Never use Clay tools. No query-objects, no ask-question-about-accounts, no Clay audiences.
3. Apollo is cross-check-only. Never the primary write path. If Landbase has a coverage gap, flag the gap rather than silently filling from Apollo.
4. Never auto-send. This skill produces files. It does not send emails, post to Slack, or create CRM engagements without explicit confirmation.
5. Deliver as a Google Sheet (when a Drive MCP is available). After writing enriched.csv, push it to Google Drive via the Drive MCP create_file tool with contentMimeType=text/csv — Drive auto-converts to a Google Sheet. Return the live https://docs.google.com/spreadsheets/d/<id>/edit URL in your summary and save it to notes.md. A bookmarkable live link is the expected deliverable, not just a CSV on disk. If no Drive MCP is configured, deliver the CSV and say so.

Folder convention

Every prospect run gets its own folder under the user’s working directory:

prospects/YYYY-MM-DD_<slug>/

Inside:

• brief.md — the original ask, the criteria, source-of-truth notes
• query.json or query.md — the Landbase API query / parameters used
• raw.jsonl or raw.csv — Landbase API output (full pull, never truncated)
• enriched.csv — final ranked / scored output (the deliverable)
• outreach/ — personalization drafts, if requested (drafts only, never sent)
• notes.md — anything notable: gaps, recommendations, the live Sheet URL

Workflow

1. Capture the ask — what’s the ICP? Industry, size, geo, role, signals, exclusions? If any are fuzzy, ask the user before querying.
2. Create the folder prospects/YYYY-MM-DD_<slug>/ and write brief.md summarizing the ask.
3. Sharpen a fuzzy ICP first. If the ask is vague (“CEOs in tech,” “AI companies”), tighten it into 2–3 concrete, queryable sub-segments before pulling — confirm with the user.
4. Compose the Landbase query. Use the documented endpoints (see docs.landbase.com). Many enriched fields are available — select what the ask needs, not everything.
5. Run the query and save raw output to raw.jsonl (or .csv). Never truncate raw data.
6. Enrich / dedupe through Landbase. If a field is missing, flag it in notes.md as a Landbase coverage gap. Do not fall back to Apollo unless the user explicitly approves.
   • AI qualification (recommended): Before ranking, qualify the enriched dataset against the ICP criteria. Translate the ask into yes/no questions and run:

     landbase-cli workflow qualify ds-ENRICHED \
       --qualification-prompts="<yes/no question from ICP criteria>" \
       --context-cols=company_name,website,industry,description,employees_count \
       --wait
     

     In an interactive session, trigger without --wait, capture the workflow_run_id, and follow the workflow-monitor skill to narrate live progress — a --wait subprocess can’t stream its progress line into the chat. Keep --wait for an unattended run where only the final result matters. Use the qualify-leads skill for the full guided workflow. Widening first then qualifying yields 50%+ more results than filtering by keywords alone.

7. Score and rank. Default dimensions (override per ask): ICP fit, deal-size signal, intent signals, role seniority, and network overlap (cross-reference a LinkedIn connections export for warm intros if the user provides one).
8. Output enriched.csv with the top N ranked. Suggested columns: name, company, title, why-they-fit, suggested hook (one line, optional), warm-intro-available (Y/N).
9. Push to Google Sheets (if a Drive MCP is available). Read enriched.csv and call the Drive MCP create_file with title="<human title> (YYYY-MM-DD)", textContent=<csv body>, contentMimeType=text/csv. The response contains an id; the live URL is https://docs.google.com/spreadsheets/d/<id>/edit. Save it to notes.md under a ## Live deliverable header.
10. Write notes.md with: the live Sheet URL (if created), what worked, what’s missing in Landbase, and a suggested next iteration.
11. Surface to the user — one paragraph: the run summary, the live Sheet URL (prominent, if created), the top 5 names, the file paths, and any flags. Don’t dump the full list in chat — reference the Sheet/file.

Personalization (optional sub-mode)

If the user asks for “personalized outreach” or “hooks per row”:

1. Per row, pull additional signal: recent LinkedIn activity, recent company news (e.g. via web search).
2. Generate a one-sentence hook tailored to the row.
3. Drafts only. Write to outreach/<contact-slug>.md. Never send.
4. Match the sender’s voice if known — terse, hook-led, no platitudes; keep it concise.

Red Flags — STOP

• User wants info on one specific company → use match-lookup instead
• ICP is fuzzy and undefined → run icp-capture first before building the list
• User wants long-form content or posts → out of scope; decline
• User wants to sync to CRM → separate explicit operation; ask for confirmation first

Output discipline

• Date-stamped folder. No exceptions.
• Markdown for human-readable artifacts; CSV/JSONL for data.
• Never dump the raw export into the root of prospects/ — always use the run folder.
• Note Landbase coverage gaps explicitly — these are useful product feedback.

When in doubt

Ask the user: which ICP, what size of list (10, 100, 1000?), what’s the use (outbound, partnership, M&A, content?), and the deadline. One round-trip of clarification beats a 1000-row list that misses the ask.