You design the outbound plan.
I already run the machine.

How I do it

• leads.py takes niche, city and count, shows the search cost in a dry-run first, then searches.
• vinder.py / scanner.py read company sites and public search hits.
• Known contacts are removed against local archives and Attio before anything else happens.
• A second route pulls lists from Apollo: apollo_attio.py ophalen → voorstel → schrijf.
• Night run nachtleads.sh works in batches of 15.

For Hebno

• ICP profile file per segment: research & insight consultancies first, then strategy, creative and advisory.
• Fit classes the AI judge picks from: fits · solo · agency-not-consultancy · supplier · other, each with a confidence.
• Output: leads.jsonl + leads.csv per round, with the source URL for every field.

How I do it

• verrijk.py: Claude extracts who decides now and which sentence on their site is a reason to talk, with the literal quote as evidence.
• Names and quotes are re-checked by code against the source. Guessed email addresses are labelled uncertain, never presented as verified.
• verrijking_attio.py writes company + person + research note into Attio. Dry by default, --echt writes.

Fields I'd enrich for you

• practice_areas, client_work_type (research, strategy, creative…)
• ai_stance: how they talk about AI today, with quote + URL
• decision_maker, role, email_source, hook_quote, confidence

How I do it

• Claude writes a short mail from the recipient's problem with one simple question, using only facts from the enrichment. No invented personalisation.
• intentie.py picks the angle; the sending step refuses a record without mail_bron (where the address was found).
• Rule gate: a mail under the minimum word count or without subject is rejected as empty letter; a source that 404s is rejected as dead source.

For Hebno

• The founders write the sequences and copy; I encode them as templates + rules (banned claims, length, tone, one CTA).
• A blind checker model reviews each draft against the rules before it can be enrolled.

Email, how I do it

• cold_verzend.py: dry-run by default, --send sends. Separate outbound domain, so the main domain is never at risk.
• Hard caps: max 3 senders, max 25 per sender, 75 per day. Warm-up ladder in opwarmstand.py can go lower, never higher.
• Send window: Tue–Thu, 08:30–10:30 local time.
• One follow-up after 7 days, in the same thread, from the same sender.
• Suppression list (addresses + domains); if it's missing, sending stops.

LinkedIn

• For my own accounts I deliberately never automate LinkedIn actions or scrape it; the system drafts touches and a human sends them.
• For you I'd wire a LinkedIn tool you approve (e.g. via Clay or a sequencer) with daily caps per seat, and the same stop rules as email.

How I do it

• Every send and every skip goes to events/cold-YYYY-MM-DD.jsonl with sender and reason.
• attio_cold_email.py sets up the cold_email list in Attio once (stages + fields, idempotent); after each accepted mail the company's entry is written.
• Dedup on domain, so no account is approached twice.
• HubSpot: I've also logged audit leads from poldermarketing.nl into HubSpot as contacts + companies.

CRM events I'd log for you

{"event":"touch_sent","channel":"email","seq":"insight-v1","step":1,
 "account":"example-consultancy.com","sender":"founder@…","ts":"…"}
{"event":"reply","sentiment":"positive","routed_to":"founder","seq_stopped":true}
{"event":"status","from":"in_sequence","to":"meeting_booked"}

How I do it

• Before every follow-up the inbox is checked: a reply, unsubscribe or bounce stops the sequence, also after mail 1.
• coldmail_log.py blocks any follow-up to someone who replied.

For Hebno Blueprint

• A small agent classifies replies: interested · not now · wrong person · unsubscribe · out of office.
• Interested → instant Slack ping to the right founder with the thread + account brief. Wrong person → referral captured as new contact.