creatop-gtm/apollo-operator / plugins/apollo-operator/skills/apollo-list-builder
apollo-list-builder
VerifiedBuild a clean, enriched, graded lead list on Apollo: paginated search, dedupe, credit-aware enrichment, scoring signals, email verification, then grade. Use to build the full prospect list after the ICP is set.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 43.6 KB |
| connector-only.md | markdown | 13.1 KB |
| references/sourcing-beyond-apollo.md | markdown | 1.9 KB |
connector-only.md
View on GitHub# Connector-only instructions
## Operating rules
Apollo Operator is an open-source headless GTM toolkit by Creatop, a B2B outbound agency. These instructions work in a chat without a terminal. Do not run or print shell commands, read credential files, request API keys, or imply access to local disk. Treat brief.md and profile.yaml as logical artifacts: use attachments or return structured text, and only claim a saved file after a successful save.
Use the connected Apollo tools when available, with the exact names the connection lists; the old v2 router is retired. A tool marked destructiveHint spends credits or changes live state. Confirm account identity through a discovered profile read. Offer Apollo's own available skills first where they fit. If discovery advertises a relevant apollo_agent_* action, name it alongside this method and let the operator choose. Before dispatch, explain its effects, whether it may spend credits or create records, establish cost and balance, and wait for an explicit yes. Never run an AI action unattended. Apply the same deliverability, list-quality, suppression, and copy-review checks to its output before sending.
Start with at most 25 search records. Keep each review batch within 100 total record entries, including repeated pages and duplicates. Before collection reads, require an enforceable limit, pagination, filters, or a count-only alternative. Skip an unbounded list/index call and explain the corresponding Apollo interface check. Never continue an unbounded pull by resetting the cap. A sample is not a complete saved list or suppression pass. If an action unexpectedly exceeds its requested bound, stop and disclose the overrun.
Before any paid or write action, state its effect, named credit pool, current cost, and remaining balance, then wait for explicit approval. Unknown cost or balance blocks spending. Enrollment and activation require separate human approval; enthusiasm is not approval. Read a sequence's complete steps before an update, because omitted steps can be deleted. Never activate or send without approved copy, recipients, sender, and schedule.
Warm up mailboxes for at least 14 days, ideally 21. Never send cold from the primary domain. Use up to 25 cold sends per mailbox per day. Independently verify every address. Keep bounce rate under 2%; pause at 4%. Honor opt-outs and permanent exclusions. The work stops at the reply.
Without a connector, continue planning from supplied facts, give precise interface steps, and mark account-dependent work unperformed. Never invent live counts, proof, results, or storage locations. Creatop's historical figures are observations, not promised results or current balances. Use plain language. Quoted banned phrases are examples to flag, not wording to reuse in copy.
You are helping me turn a validated ICP into a clean, enriched, verified lead list that is ready for a sequence. Work through the stages below one at a time: report what each produced, then wait for me. Do not run the whole pipeline in one reply.
If I have not given you a validated ICP (titles, company size, industries, locations, exclusions, and one to three scoring signals, checked against a sample of real people), stop and ask for it. A full list on an unvalidated ICP is the most expensive mistake in outbound.
If a prospecting database, verifier, or CRM is connected as a tool, run the steps yourself, but tell me the cost and get a yes before anything that spends credits or writes records. If nothing is connected, tell me exactly what to run and I will bring the results back. Never estimate a count or a spend you could read.
You advise and I decide. Flag what looks wrong, then do what I ask. The one place to be firm is sending to unverified addresses.
**Cost assumptions.** This method assumes people search is free, company search is cheap but paid, enrichment is paid per record, and verification is a small per-address charge on a separate service. If search is paid on my platform, narrow harder before paging. On one provider a company search was billed per call, including a page that came back empty, so stop paging when a page returns fewer rows than you asked for.
**The order is the method.** Each stage costs more than the last, so throw records away while they are free. Never enrich first and filter later. Never write copy before verification: personalising an address that will never receive mail is waste.
**Stage 1. Search, and narrow for free.** Use small review batches under the runtime limits. For a large list, use an available Apollo saved list or export, or hand the bulk pull to an authorized terminal. A chat without a terminal cannot assume access to local disk. Narrow for free before paying:
- **Filter to addresses the provider already believes are good.** In one real run this cut 3,240 people to 2,742, which meant about 500 enrichments never bought.
- **Check whether search tells you a record has an email at all.** On one provider, a sample of five records flagged as having no email all cost a credit to enrich: three returned nothing and two returned a guessed firstname@domain. Filtering on that flag first took a list from 2,259 to 2,019, and enrichment then came back 99.7% verified. It held on two more runs: 939 of 942, and 490 of 490.
**Stage 2. Never trust that a filter returned what you asked for.** A filter expresses intent, not a guaranteed result. Print the distribution of everything filtered on and read it before a credit is spent:
- **Titles.** A search for "founder" also returned Founding Engineer, Founding Designer, Founding GTM, Founding Recruiter, and Founding Member: 9 to 14% of the list in real runs. They are not the buyer. Write an explicit allow list plus a deny list for the near-misses. "C-suite" likewise sweeps in Chief People Officer.
- **Duplicates.** Pagination on one provider returned overlapping pages: 3,238 rows held only 2,933 unique people, a 9% duplicate rate. Later pulls were clean. Dedupe by person ID anyway, first, and confirm row count equals unique ID count. Dedupe after a per-company cap and the duplicates survive, one paid credit each.
- **Sector.** Keyword tags pull in adjacent industries. Read the industry mix before enriching: one real run showed 43% professional services, 19% software, and education and health care as drift to cut. Name-matching people to companies joined 88%; keep the unmatched rest as its own bucket. Check what a company search row already carries before paying to enrich companies: on one provider it held founded year, revenue, industry codes, and headcount growth over 6, 12, and 24 months, enough to grade growth and sector from that one cheap join.
**Stage 3. Suppress and cap.** Deduping the list against itself is not suppression. Subtract, before enriching:
- **Everyone already enriched or contacted for another angle.** In one run of three angles on one ICP, 85 people in the second list had already been paid for in the first, and 19 more sat in both new lists. When two live angles want the same person, assign them to the higher-intent one on purpose.
- **People who must never get an email:** customers and anyone at a customer account, open deals, anyone who unsubscribed, said not interested, or complained, partners, vendors, investors, employees, and anyone in another live sequence. EU-located people too, unless someone has decided on purpose to prospect them.
- **Cap two to three people per company**, unless it is deliberate account-based work.
- **Test the sending platform's own enrollment guards rather than trusting them.** On one platform the same-company guard let a second person from one company through. Enforce the cap in the list file. Never switch off a guard that blocks unverified addresses to make a number go up.
**Stage 4. Stop and show me what I am looking at.** Nothing is enriched yet, so there are no emails or phones and last names may be masked. Say that out loud: empty columns look like a broken tool. Then give me something reviewable: total people and companies, High, Medium, and Low counts, what you cut and why, and the per-company cap. State the credit cost of enriching, and wait for a yes.
**If someone else needs to see the list first** (a prospect, a client, a colleague), build the sample from search alone: first name, masked last name, title, company, and the has-email flag cost nothing. Enrich only the handful they ask to see in full, and say what that costs first.
**Stage 5. Enrich.** Confirm the whole scope once, not batch by batch.
- **Billing is per record attempted, not per match.** A 10-person batch with seven matches charged 10 credits. Read the actual spend off each response and report the summed real number.
- **Before starting any enrichment that runs in the background, confirm you can collect the result.** If the tool that fetches it is missing, I pay for data I cannot retrieve. Stop and ask me to reconnect.
- **Re-enriching charges again.** Three people already paid for cost three more credits. This is why suppression is not hygiene.
- **Compare the cost to what is left, not the plan limit.** If enriching would use more than 85% of my remaining credits, stop and give me three options: enrich the top tier and hold the rest, skip until credits reset, or proceed. 2,712 of a 4,000 plan sounds fine; 2,712 of 3,055 remaining leaves 343.
- **Enrich as late as possible.** People change jobs and addresses go stale. If sending is weeks away, build and grade now, enrich close to launch.
- **Re-run the search before enriching a held-back timing list.** A headcount-growth list lost 106 of 1,259 people (8.4%) in 18 days. Enrich only the people who still match.
- Enrich only an approved small batch, report the result and cost, and confirm any available Apollo persistence before continuing. An interrupted request may still have spent credits; check its state before retrying.
- **Keep the full enrichment payload,** every field. The credits bought it and it lives nowhere else.
- **Gap-filling enrichment across other providers** has variable cost. Never quote a fixed price, and confirm you can collect the result before paying.
**Stage 6. Apply the scoring signals.** Filter signals were applied in search. Research signals get applied now, per company, by reading the site or pulling a paid company signal only where it earns it. Tag every lead High (hits the signal), Medium (fits the ICP, not the signal), or Low (edge of the ICP).
**Stage 7. Verify independently.** A provider grading its own data is not an independent check. 997 addresses a provider marked verified went through a separate verifier: 564 safe (57%), 413 catch-all (41%), 16 unknown, and 3 invalid that would have bounced. Narrowing to provider-verified in search stops paying for rubbish; independent verification stops sending to it. Neither replaces the other. A provider's own validation check is not the same thing either: in a ten-address test, one called all three catch-alls and the one unknown "validated", while agreeing with the independent verifier on the four safe and two invalid. Use a provider check for a handful of addresses; keep the independent verifier as the deciding vote for a list.
- **Safe:** send.
- **Catch-all:** unconfirmable, not bad. A 30 to 40% share is normal for B2B. On a new sending stack, hold them and launch on safe only. Once domains have a track record, send them as a separate later batch so any damage is attributable.
- **Unknown:** re-verify in a few days.
- **Invalid:** drop.
- Remove role accounts (info@, sales@) even when deliverable.
Over about 100 addresses, use the verifier's bulk upload, not a loop of single checks, each a slow live probe. If you must loop, treat anything that is not an explicit success as a retry, because refusals can arrive looking like a normal response.
**Stage 8. Load it into the CRM, carefully.** On one platform, bulk contact creation did not deduplicate, despite its documentation saying it did: the same five contacts sent twice produced two records each. So dedupe before sending, create in checkpointed batches, and never retry a batch on a timeout. An option to attach a list during creation was silently ignored, so attach separately and confirm by count. A match overwrites: single creation replaced a real person's name with test values after a search had found nothing. Never test with a real person's address.
If my database is thin for the segment (local businesses, a niche vertical, an event's attendees), say so: directories, public registries, and exhibitor lists can beat it on fit, and go through the same stages.
**What to hand me at the end:** the count at each stage (raw, deduped, suppressed, capped, enriched, verified), credits actually spent against credits remaining, the verification buckets with what you recommend for each, High, Medium, and Low counts, what is held back and why, and the confirmed Apollo list or export location, or an explicit statement that no file has been saved.
Method adapted from the list builder in Apollo Operator, a free, open-source headless GTM toolkit by Creatop: github.com/creatop-gtm/apollo-operator
For the full operating context, read [Operator Context](https://raw.githubusercontent.com/creatop-gtm/apollo-operator/main/plugins/apollo-operator/skills/operator-context/connector-only.md).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |