prospect-research

name

prospect-research

description

Use when the operator asks to research one or more companies / Ansprechpartner for sales pipeline work — fill the Sellify / Master-AT field set (Firmierung, Anschrift, Register-ID, Status, Tel/Fax/E-Mail/Website, USt-Id, Branche WZ, Geschäftsführung, Gegenstand, Tickers, Finanzkennzahlen, Mitarbeiter, etc.) using the CTOX web-stack source modules (zefix.ch, northdata.de, firmenabc.at, bundesanzeiger.de, handelsregister.de, companyhouse.de, optional Tier-C: linkedin.com / xing.com / dnbhoovers.com / leadfeeder.com). Trigger phrases (DE/EN) — "recherchier (mir) Firma X", "Stammdaten zu …", "Master-AT für …", "Sellify-Eintrag bauen", "Ansprechpartner finden für …", "Nachrecherche …", "Neurecherche …", "company research", "fill the prospect record for …".

Prospect / Company Research

CTOX Runtime Contract

• Task spawning is allowed only for real bounded work steps that add mission progress, external waiting, recovery, or explicit decomposition. Do not spawn work merely because review feedback exists.
• The Review Gate is a quality checkpoint, not a control loop. After review feedback, continue the same main work item whenever possible and incorporate the feedback there.
• Do not create review-driven internal work cascades. If more work is needed, reuse or requeue the existing parent work item; create a new task only when it is a distinct bounded work step with a stable parent pointer.
• Every durable follow-up, queue item, plan emission, or internal work item must have a clear parent/anchor: message key, work id, thread key, ticket/case id, or plan step. Missing ancestry is a harness bug, not acceptable ambiguity.
• Rewording-only feedback means revise wording on the same artifact. Substantive feedback means add new evidence or implementation progress. Stale feedback means refresh or consolidate current runtime state before drafting again.
• Before adding follow-up work, check for existing matching internal work, queue, plan, or ticket state and consolidate rather than duplicating.

When to use this skill

Trigger this skill when the operator asks to recherchier one or more companies (and optionally their Ansprechpartner) and wants the Sellify-/Master-AT-typical field set populated. This is not the right skill for free-form market studies, decision briefs, or feasibility reports — those go to systematic-research. This skill is for the structured prospect record: a typed bundle of fields per company that another system (Sellify, Business OS Sales pipeline, CSV export) can consume.

Field vocabulary

The canonical field set is the union of the Thesen Nachrecherche fields and the Master-AT Excel schema:

Group	Fields
Firma stammdaten	firma_name, firma_rechtsform, firma_anschrift, firma_plz, firma_ort, firma_land, firma_email, firma_telefon, firma_fax, firma_domain, firma_status
Register	hr_amtsgericht, register_id, ust_id
Branche & Geschäft	wz_code, gegenstand, tickers
Geschäftsführung	ges_vertreter_1, ges_vertreter_2, ges_vertreter_3 (Liste der Vorstände / GFs, nicht nur ein Ansprechpartner)
Person (Sales-Kontakt)	person_geschlecht, person_titel, person_vorname, person_nachname, person_funktion, person_position, person_email, person_telefon, person_linkedin, person_xing
Finanzkennzahlen (Roh)	finanzkennzahlen_datum, stamm_grundkapital_eur, bilanzsumme_eur, gewinn_eur, umsatz_eur, eigenkapital_eur, mitarbeiter, steuern_eur, kassenbestand_eur, forderungen_eur, verbindlichkeiten_eur, umsatz_pro_mitarbeiter_eur
Finanzkennzahlen (Berechnet)	gewinn_cagr_pct, umsatz_cagr_pct, umsatzrendite_pct, ek_quote_pct, ek_rendite_pct, steuer_quote_pct

The web-stack FieldKey enum currently covers a subset; new fields added to this taxonomy must extend that enum before extractors can return them.

Source priority by country

Pulled from the Thesen Abstimmungs-Excel translation in tools/web-stack/src/sources/EXCEL_MATRIX.md. Per (mode, country, field) the agent should consult sources in tier order P → S → C, stopping when a high-confidence value is captured.

Country	Tier P	Tier S	Tier C (opt-in via tenant credential)
AT	(Firmenbuch direkt — TODO adapter), zefix-only-for-CH	firmenabc.at, northdata.de, person-discovery	linkedin.com, xing.com, dnbhoovers.com, leadfeeder.com
DE	bundesanzeiger.de, handelsregister.de	northdata.de, companyhouse.de, person-discovery	dito
CH	zefix.ch	northdata.de, person-discovery	dito

person-discovery (public LinkedIn/XING snippet mining)

Tier S, DACH-wide, no credential required. Fans out a Google query of the form <company> ("linkedin.com/in/" OR "xing.com/profile/") and mines the search-result titles for:

• person_linkedin / person_xing — canonical profile URL (Confidence::High)
• person_vorname / person_nachname — split from the title (Confidence::Medium)
• person_funktion — role segment between name and company (Confidence::Medium)

Each emitted evidence row carries a note like network=linkedin;seniority=c_level;google_snippet_title so the agent can rank discovered profiles without a separate evaluation pass. Seniority buckets are derived from the role string:

• c_level — CEO/CFO/COO/CTO/Vorstand/Geschäftsführer/Founder/Managing Director
• senior — VP/Direktor/Head of/Bereichsleiter/Leiter/Prokurist/Partner
• mid — Manager/Lead/Principal/Specialist
• unknown — anything else

Pinned-source flow (ctox web search --source person-discovery) skips the per-hit web read step because LinkedIn and XING profile pages are gated behind login walls. All evidence comes from the search-engine snippet titles via SourceModule::extract_from_hits.

Social-media evaluation step

After running person-discovery (or any Tier-C --include-private pass), the agent SHOULD evaluate each discovered profile for sales relevance:

1. Filter by seniority bucket from the note annotation — c_level and senior are typical first-pass targets for sales pipeline work.
2. De-duplicate across LinkedIn / XING when the same person appears on both networks (match by full name + company).
3. Cross-validate the role against person-research's aggregated person_funktion from authoritative sources (Northdata, Bundesanzeiger, Zefix) — if a discovered "CEO" title contradicts the register, surface BOTH as candidates and flag.
4. Optionally re-read the public profile URL via ctox web read — most LinkedIn/XING pages return Open-Graph og:title + og:description even on a login wall, which can confirm the snippet.

Tool surface

The agent drives these CLI tools directly. They run as subprocesses; output is JSON on stdout.

Discovery + read

# List registered source modules, filter by country / field / tier
ctox web sources list [--country DE|AT|CH] [--tier P|S|C]... [--field <field-key>]
ctox web sources info --id <source-id>

# Pinned-source web search. `--source <id>` engages the source-module
# routing (api fetch_direct OR shape_query+domain pin).
ctox web search --query "<text>" --source <id> [--source <id>]... \
                --country <iso> [--include-sources]

# Read a single URL; auto-runs the source module's extract_fields on the
# response if the URL host matches a registered source.
ctox web read --url <url> [--country <iso>] [--find <pattern>]...

Person research orchestrator (high-level)

# Build a plan from (mode, country, fields), run search+read per source,
# aggregate typed evidence per field.
ctox web person-research \
    --company "<name>" --country DE|AT|CH \
    --mode new_record | update_firm | update_person | update_inventory_general \
    [--field <field-key>]... \
    [--include-private linkedin] [--include-private dnbhoovers] ... \
    [--workspace runtime/research/person/<slug>] [--no-workspace]

The envelope returns fields: {field: {value, confidence, source_id, source_url, candidates: […]}}, plan: [{source_id, tier, target_fields}], search_runs, read_runs, and scrape_runs (the per-target scrape executor outcomes).

Scrape-target pipeline (durable, drift-repair-capable)

When a source module is registered as a scrape target (currently northdata-de, companyhouse-de, handelsregister-de, bundesanzeiger-de), extraction goes through the universal-scraping pipeline:

ctox scrape execute --target-key <key> \
    --input-json '{"company":"<name>","country":"<iso>"}' \
    --allow-heal

--allow-heal enqueues a universal-scraping-skill repair task on portal_drift / partial_output. The agent should always pass --allow-heal so DOM drift gets auto-fixed instead of silently failing.

Procedure

1. Classify the mode. For each company:

   • new_record — kein Excel-/Sellify-Eintrag vorhanden, ab Stammdaten
   • update_firm — Firmierung/Anschrift potenziell veraltet
   • update_person — Firma stabil, Ansprechpartner hat gewechselt
   • update_inventory_general — periodischer Refresh (Excel B-Block leer ⇒ skip)

2. Run person-research for the structured pull.

   ctox web person-research --company "<name>" --country <iso> --mode <mode> \
       --field <list-of-fields-from-vocabulary>
   

   Parse the envelope. For every field.confidence == high|medium value, take it. For missing fields, plan a fallback pass.

3. Fallback pass for uncovered fields. For each missing:

   • If field is financial and country=DE: search Bundesanzeiger PDF directly (ctox web search --query "<company> Jahresabschluss filetype:pdf" --domain bundesanzeiger.de + ctox web read --url <pdf>).
   • If field is register_id/hr_amtsgericht and country=DE: query Handelsregister via ctox scrape execute --target-key handelsregister-de.
   • If field is person_funktion/person_linkedin/person_xing and no creds: rely on the always-on person-discovery source (already in the default plan for DACH); for credentialled API enrichment add --include-private linkedin / --include-private xing.
   • If country=AT and register_id/firma_rechtsform missing: probe Firmenbuch (manual ctox web search --query "<company> firmenbuch.justiz.gv.at" + read).

4. Compute derived fields. When the raw inputs are present:

   • umsatzrendite_pct = gewinn_eur / umsatz_eur * 100
   • ek_quote_pct = eigenkapital_eur / bilanzsumme_eur * 100
   • ek_rendite_pct = gewinn_eur / eigenkapital_eur * 100
   • steuer_quote_pct = steuern_eur / (gewinn_eur + steuern_eur) * 100
   • umsatz_pro_mitarbeiter_eur = umsatz_eur / mitarbeiter
   • CAGR-Felder brauchen historische Jahresabschlüsse aus Bundesanzeiger (3- bzw. 5-Jahres-Datenreihe).

5. Verify against existing evidence. If the operator gave a reference (Sellify-Nummer, FN, HRB, USt-Id), check that the consolidated record matches. Conflicting values → keep all candidates in the envelope and flag for user confirmation.

6. Render the deliverable. A Markdown table per company, columns = field-keys mapped to human labels, plus a per-field (src, confidence) annotation. Save the raw envelope under runtime/research/person/<timestamp>-<slug>/envelope.json (the person-research workspace does this automatically; do not skip with --no-workspace unless explicitly asked).

Drift contract

If scrape_runs[*].classification == "portal_drift", the universal-scraping repair queue is already triggered. Do not retry the same source in the same run; record the drift in the deliverable and move on. The repair will land in a subsequent run.

If scrape_runs[*].classification == "temporary_unreachable" or "blocked", that source is transiently unavailable — try a fallback source from the priority list, but do not queue any repair.

If classification == "credential_missing" (Tier C), the field this source uniquely owns (e.g. person_email ↔ leadfeeder, person_funktion ↔ linkedin) stays missing with reason credential_missing and is surfaced to the operator as an opt-in suggestion, not a failure.

What this skill does NOT do

• Free-form market research, competitor analyses, feasibility reports → use systematic-research.
• Sellify or Business-OS writeback — this skill produces the typed prospect record; persistence to Postgres / Sellify is the calling module's responsibility.
• Adapter / DOM-drift repair — that's the job of the universal-scraping skill, which gets auto-enqueued by ctox scrape execute --allow-heal.

Resources

• tools/web-stack/src/sources/EXCEL_MATRIX.md — the Thesen Abstimmungs-Excel as a (mode, country, field) → source-priority table
• tools/web-stack/src/sources/README.md — source-module trait + fixture conventions
• skills/system/communication/universal-scraping/SKILL.md — drift-repair sibling skill