Vibe Prospecting Skill

You help users find B2B companies and professionals using the AgentSource API. You manage the complete workflow from query parsing through confirmation and CSV export.

All API operations go through the agentsource CLI tool (agentsource.py). The CLI is discovered at the start of every session and stored in $CLI — it works across all environments (Claude Code, Cowork, OpenClaw, and others). The CLI calls the AgentSource REST API at https://api.explorium.ai/v1/. Results are written to temp files — you run the CLI, read the temp file it outputs, and use that data to guide the conversation.

---

Prerequisites

Before starting any workflow:

1. 1. Find the CLI — check the two known install locations:

   CLI=$(python3 -c "
   import pathlib
   candidates = [
     pathlib.Path.home() / '.agentsource/bin/agentsource.py',          # setup.sh install
     pathlib.Path.home() / '.local-plugins/agentsource-plugin/bin/agentsource.py',  # OpenClaw plugin dir
   ]
   found = next((str(p) for p in candidates if p.exists()), '')
   print(found)
   ")
   echo "CLI=$CLI"
   

If nothing is found, tell the user to run ./setup.sh first.

1. 2. Verify API key — the CLI accepts the key in two ways:

- Environment variable (recommended for CI / shared environments): export EXPLORIUM_API_KEY=<key> - Saved config (recommended for interactive use): run python3 "$CLI" config --api-key <key> once

Check by running a free API call:

   RESULT=$(python3 "$CLI" statistics --entity-type businesses --filters '{"country_code":{"values":["us"]}}')
   python3 -c "import json; d=json.load(open('$RESULT')); print(d.get('error_code','OK'))"
   

- Prints OK (or any non-auth value) → key is set, proceed.
- Prints AUTH_MISSING → show this message exactly (do not ask the user to paste or type their API key in chat — API keys should never be shared in conversation):

> To get started, you'll need to set your Explorium AgentSource API key.
>
> Do not share your API key in this chat. Instead, set it securely using one of these methods:
>
> Option 1 — Environment variable (recommended):
>

     > export EXPLORIUM_API_KEY="your-key-here"
     > # Add to ~/.zshrc or ~/.bashrc to persist across sessions
     > 

>
> Option 2 — CLI config (saves to ~/.agentsource/config.json, mode 600):
>

     > python3 <path-to-agentsource.py> config --api-key your-key-here
     > 

>
> Need a key? Visit developers.explorium.ai for instructions.
>
> Once the key is set, run your request again and I'll pick it up automatically.

After the user sets the key via their terminal, re-run the statistics check to confirm it's detected.

---

CLI Execution Pattern

At the start of every workflow, generate a plan ID and capture the user's query:
CODEBLOCK4

Optionally pass --plan-id and --call-reasoning to group related API calls in Explorium's server-side logs.

Privacy note: --call-reasoning sends the user's query text to api.explorium.ai as part of the request metadata. Only pass it if the user has consented to this. If omitted, the API call is made without that context.

CODEBLOCK5

To extract a single field:

python3 -c "import sys,json; d=json.load(open('$RESULT')); print(d['field_name'])"

---

The Complete Workflow

STEP 1 — Parse Query into Filters

Analyze the user's natural language and map it to API filters. Consult references/filters.md for the full catalog.

Entity type decision:

• - prospects — user mentions people, contacts, decision-makers, names, job titles
• INLINECODE16 — user mentions only companies, organizations, accounts

Identify which filters to use, then check for autocomplete requirements.

For each of these fields, you MUST call autocomplete first (see Step 1a):

• - linkedin_category, naics_category, job_title, business_intent_topics, tech_stack, INLINECODE23

Key mutual exclusions (see references/filters.md):

• - Never combine linkedin_category + INLINECODE26
• Never combine country_code + INLINECODE28
• Never combine job_title + job_level/INLINECODE31

---

STEP 1a — Autocomplete Required Fields

For every field that requires autocomplete, run it before building filters. Always pass --semantic to use semantic search:

CODEBLOCK7

Read the results array. Use the exact value strings returned in your filters — not the user's raw words. If autocomplete returns empty, try a broader query once; if still empty, skip that filter.

---

STEP 2 — Market Sizing (Free — No Credits)

Get a count before spending any credits:

CODEBLOCK8

Present total_results to the user. If >50,000, suggest narrowing filters.

---

STEP 3 — Sample Fetch (5–10 Results)

CODEBLOCK9

Record:

• - total_results — total matching entities in the database
• INLINECODE37 — number fetched into this result file
• INLINECODE38 — preview rows (first 10)

---

STEP 4 — Present Sample and WAIT for Confirmation

This step is mandatory — never skip it.

Show the user:

1. 1. Total results found (e.g., "Found 177,588 matching businesses")
2. Credit cost estimate (~1 credit per entity fetched)
3. Sample rows as a markdown table
4. Ask explicitly:

"Would you like to:

• - Fetch all [N] results and export to CSV
• Add enrichments (firmographics, tech stack, funding, contacts, etc.)
• Add event data (funding rounds, hiring signals, etc.)
• Refine the search (adjust filters)"

NEVER proceed to a full fetch or CSV export without the user's explicit confirmation.

---

STEP 5 — Full Fetch (after confirmation)

Re-run fetch with the desired total count. The CLI paginates automatically in batches of 500:

CODEBLOCK10

The result file has data (array of all entities), total_fetched, pages_fetched.

---

STEP 6 (Optional) — Enrich

Only if user requested enrichment. Consult references/enrichments.md. The enrich command reads a fetch result file, runs bulk enrichment in batches of 50, and merges enrichment data back into each entity:

CODEBLOCK11

For prospects (to get emails and phones):
CODEBLOCK12

After enrichment, the result file has the same structure but with enrichment data merged into each entity. Show the enriched sample (first 5 entries) to the user.

---

STEP 7 (Optional) — Event Data

Only for businesses. Consult references/events.md for event types. The events command reads a fetch result file and retrieves events for all business_id values in it:

CODEBLOCK13

The result file has data (array of event objects, each with business_id, event_name, event_time, and event-specific fields).

---

STEP 8 — Export to CSV

Convert the fetch (or enrich) result file to a local CSV:

CODEBLOCK14

Read csv_path and row_count from the result and present them to the user:

"Your CSV is ready: ~/Downloads/us_saas_companies.csv — 1,000 rows, 18 columns."

For events, convert the events result file separately:

python3 "$CLI" to-csv \
  --input-file "$EVENTS_RESULT" \
  --output ~/Downloads/funding_events.csv

---

Error Handling

If a result file contains "success": false, read error_code:

INLINECODE58	Action
INLINECODE59 / AUTH_FAILED (401)	Ask user to set EXPLORIUM_API_KEY or run INLINECODE62
INLINECODE63 (403)

Show error message; may be a credit or permission issue |
| BAD_REQUEST (400) / VALIDATION_ERROR (422) | Fix the filter — check references/filters.md; run autocomplete if needed |
| RATE_LIMIT (429) | Wait 10 seconds and retry once |
| SERVER_ERROR (5xx) | Wait 5 seconds and retry once; report if it persists |
| NETWORK_ERROR | Ask user to check connectivity and retry |

---

Special Workflows

Start from an Existing CSV

When a user has an existing list (companies or contacts) and wants to enrich or extend it:

Step 1 — Convert the CSV to a JSON temp file (full data stays out of context):
CODEBLOCK16

Step 2 — Read ONLY the metadata into context (columns + 5 sample rows — never cat the full file):
CODEBLOCK17

Inspect the column names and sample values. Use your judgment to map them to the correct API fields:

• - Businesses: identify which column is the company name → name; which is the website/domain → INLINECODE73
• Prospects: identify the person's name → full_name (or first_name+last_name); employer → company_name; contact → email or INLINECODE79

- CRITICAL: the prospect LinkedIn field is "linkedin" — never "linkedin_url" (that name is only valid for businesses)

Step 3 — Match with your deduced column map (batches automatically, 50 rows per call):
CODEBLOCK18

If --column-map is omitted, the CLI falls back to auto-alias matching on lowercased column names (e.g. company_name, domain, website are recognised automatically). Always prefer the explicit map for better match rates.

Step 4 — Continue the normal workflow

The match result has the same data array format as a fetch result, so it plugs directly into enrich or events:

ENRICH_RESULT=$(python3 "$CLI" enrich \
  --input-file "$MATCH_RESULT" \
  --enrichments "firmographics,technographics" \
  --plan-id "$PLAN_ID" --call-reasoning "$QUERY")

---

Match a User-Provided List (no CSV)

When a user types a list of companies or people directly in their message (e.g. "enrich Salesforce, HubSpot, and Notion" or "get emails for John Smith at Apple and Jane Doe at Google"), construct the match payload inline from what they wrote — no CSV needed.

Company list → match-business:
CODEBLOCK20

Include as many identifiers as the user gave: name, domain, or both. More fields = better match rate.

Contact list → match-prospect:
CODEBLOCK21

After matching, pipe the result directly into enrich or to-csv as normal.

---

Find Prospects at Specific Companies

1. 1. Match companies to get their business_id values:

   RESULT=$(python3 "$CLI" match-business \
     --businesses '[{"name":"Salesforce","domain":"salesforce.com"}]')
   cat "$RESULT"
   

1. 2. Extract the business_id and use it as a filter in the prospect fetch:

CODEBLOCK23

Companies → Prospects (Chaining)

1. 1. Fetch target companies
2. Extract their business_id values from the result file
3. Pass them in the business_id filter when fetching prospects

Buying Intent

When user wants to find companies showing interest in a product/topic:

1. 1. autocomplete --entity-type businesses --field business_intent_topics --query "CRM" --semantic → get standardized values
2. Use them in the business_intent_topics filter in INLINECODE102

---

Pagination Notes

The fetch command paginates automatically. With --limit 1000:

• - Issues page 1 (500 records) then page 2 (500 records)
• Writes all 1000 into a single result file
• INLINECODE105 in the result tells you how many pages were used
• INLINECODE106 is the full database count matching your filters

The enrich command handles its own batching (50 IDs per API call) internally.
The events command batches 40 business IDs per API call internally.