Prospect Discovery

Prospect Discovery

Overview

Create a customer discovery strategy before research and scoring. The core rule is to use approved business sources, record where each lead came from, and pass candidates to company-research and prospect-scoring for evidence review.

When to Use

Use this skill when the user asks how to find overseas customers, build a prospect list, plan search keywords, use a collection API, review public directories, or prepare leads for outreach.

Use it for company-level discovery planning and evidence collection before outreach drafting.

Inputs

• Product context from product-loader
• Market rules from MARKET.yaml
• Discovery settings from DISCOVERY.yaml
• Target countries or regions
• Approved collection API details, if provided by the user

Outputs

{
  "collection_api_status": "configured|none|needs_user_input",
  "keyword_strategy": [],
  "target_regions": [],
  "exclude_terms": [],
  "allowed_sources": [],
  "candidate_fields": [],
  "source_status": "verified|search_tasks_only|source_unavailable",
  "source_evidence_rules": [],
  "risk_notes": [],
  "next_steps": ["company-research", "prospect-scoring"]
}

Procedure

1. Read DISCOVERY.yaml before asking about collection tooling.
2. If collection_api.provider is missing, ask the user once for the collection API. If there is no API, write none into DISCOVERY.yaml when file editing is available.
3. If the user provides an approved API after setup, write its provider, endpoint, method, API key environment variable, auth header, query parameters, response mapping, pagination, rate limit, and retry policy into DISCOVERY.yaml; keep raw secrets in environment variables.
4. Read scraping.engine. Use scrapling-fetcher only for crawling company pages after a candidate is already known. Do not use browser navigation as the default customer-discovery path.
5. Run python tools/collect_prospects.py --discovery --product --output-dir when file output is needed.
6. If the user only names a product or SKU, pass --product-query or --sku instead of rewriting PRODUCT.yaml.
7. If the agent has already identified source pages during the task, pass them at runtime with --source-url instead of editing DISCOVERY.yaml.
8. If discovery_mode is native_scrapling_spider, scrapling_spider.enabled is true, or runtime --source-url values are provided, tools/collect_prospects.py routes to tools/scrapling_spider_runner.py, uses Scrapling's native Spider, and writes crawl_report.json.
9. If the agent has MCP access configured, call tools/scrapling_mcp_server.py through the collect_prospects MCP tool and pass runtime source_urls for the same native Spider workflow.
10. When neither a configured API, configured scrapling_spider.source_urls, nor runtime source URLs are available, create search tasks for Google search results, trade show websites, industry directories, prospect company websites, LinkedIn public summaries, and B2B platform public pages.
11. If the tool outputs only prospect_search_tasks.csv, stop and report source_status: "search_tasks_only"; ask for a collection API, a user-provided list, or permission to process a specific approved source.
12. Do not manually browse arbitrary B2B platforms after search tasks are generated.
13. Do not return a numbered customer list unless it comes from prospects.raw.csv, a user-provided prospect file, or fetched company pages with source URLs.
14. If sources cannot be accessed or provide no usable company records, return source_status: "source_unavailable" and do not fill the gap with industry knowledge.
15. With a configured API, produce prospects.raw.csv and prospects.raw.json.
16. Build keyword groups from product names, HS codes, applications, buyer types, target regions, and channel terms.
17. Add exclude terms for jobs, consumer reviews, unrelated retail-only pages, marketplaces without seller websites, and irrelevant industries.
18. Keep each candidate tied to a source URL, company-level signal, and reviewable evidence summary.
19. When discovery finds a company name or company link, visit the official website or source-linked company page next and run official website contact search before finalizing the row.
20. During official website contact search, check homepage, contact, about, team, catalog, and product pages for visible email and phone values.
21. For each candidate, record company name, website, country, business type, source URL, evidence summary, risk notes, contact email, contact phone, email result, and phone result.
22. If email or phone is not found during discovery, write contact_email: "没有", contact_phone: "没有", email_result: "没有", or phone_result: "没有" instead of omitting the fields.
23. Respect output_formats in DISCOVERY.yaml or runtime --formats when exporting prospects.raw or prospect_search_tasks.
24. Send candidates with evidence to company-research; send researched prospects to prospect-scoring.

Verification

• DISCOVERY.yaml is checked before asking for collection API details.
• Missing API is recorded as none instead of repeatedly asking.
• Search tasks are written when no collection API is configured.
• Search tasks are not treated as customer results.
• Browser navigation is not used to continue discovery after search-task output.
• Native Scrapling spider mode accepts configured scrapling_spider.source_urls or runtime --source-url values, uses tools/scrapling_spider_runner.py, and writes crawl_report.json.
• MCP mode uses tools/scrapling_mcp_server.py, accepts runtime source_urls, and returns the same output paths as the CLI runner.
• A numbered customer list requires prospects.raw.csv, a user-provided prospect file, or source URLs from fetched company pages.
• source_unavailable is used when sources do not return usable company records.
• API collection writes prospects.raw.csv and prospects.raw.json.
• A company name or company link triggers official website contact search before the candidate is finalized.
• prospects.raw.csv always includes contact_email, contact_phone, email_result, and phone_result; missing contact values are written as 没有.
• output_formats or runtime --formats controls whether the tool writes CSV, JSON, XLSX, or a mix of them.
• Configured APIs include response mapping, pagination, rate limit, and retry policy.
• Scrapling is the default scraping backend, and browser modes are enabled only when their dependencies are installed.
• Product or SKU selection can be passed through --product-query or --sku.
• Every candidate includes a public source URL and evidence summary.
• Every candidate depends on reviewable company-level evidence.
• Next steps include company-research before prospect-scoring.

Common Mistakes

| Mistake | Fix | | --- | --- | | Treating search results as verified buyers | Pass candidates to research before scoring | | Filling a requested customer count from industry knowledge | Return source_unavailable or search_tasks_only | | Saving raw API keys in project files | Save only the environment variable name | | Keeping weak source records | Store source URL, date, and evidence summary | | Mixing company and contact evidence | Keep discovery focused on company-level signals |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Xuxchloris
• Source: Xuxchloris/hermes-export-skills
• License: MIT

Install and usage instructions live in the source repository linked above.