The problem this solves
Every marketer who has commissioned a prospect list knows the failure mode: you describe the ideal customer, someone or something goes away for a day, and a CSV of eight hundred rows comes back where row three is already wrong. The criteria were interpreted, the interpretation drifted, and now the errors are baked through the whole file. Fixing it means either a painful manual scrub or a full rebuild, and either way the list arrives late and half-trusted.
The waste compounds downstream. Bad rows do their real damage after import: sequences fire at the wrong titles, reply rates sag, and your CRM accumulates records that someone eventually has to deduplicate and purge. The cost of a misread criterion is never just the list. It is the campaign built on the list, and the cleanup that follows the campaign, long after anyone remembers which assumption produced the bad rows in the first place.
The fix is embarrassingly simple and almost never operationalized: inspect a sample before committing to the run. Ten rows are enough to catch a wrong industry read, a seniority miss, or a geography leak. What has been missing is a worker that treats the sample as a checkpoint by default - one that shows its interpretation, takes your corrections in plain language, and only then spends the effort on the full build.
How the mission runs
- Define the criteria in conversation. You describe the target the way you would to a colleague: the segment, the titles, the signals that matter. The agent turns that into explicit criteria, using Intent Data to anchor the in-market signal, and reflects the interpretation to you so vague phrases become checkable definitions before anything runs.
- Ten sample rows land in the thread. The agent builds exactly ten rows and posts them where you are talking: company, person, title, and a verified address from Email Finder, plus why each matched. This is the checkpoint. The full run is explicitly held until you react, and the sample costs a fraction of the whole build.
- Correct course in plain language. Reply the way you would to a researcher: too small, wrong vertical on row four, more operations titles and fewer finance ones. The agent revises the criteria, shows a fresh sample if the change was material, and confirms the final definition. Nothing about this loop requires forms or query syntax.
- Your yes releases the full run. On explicit approval in the thread, the agent scales the approved criteria to the full 200, applying the same matching bar the sample passed. Because the run cost is projected up front and only spends after your yes, an expensive build never fires on a misunderstanding.
- CSV delivered, approved contacts into HubSpot. The finished list arrives as a CSV in the thread for your records, and the approved contacts load into HubSpot with fields mapped and duplicates against existing records skipped. Writes are additive: the agent creates and updates the records you approved and touches nothing else in your portal.
The prompt
This is the exact objective the agent receives. Swap the obvious placeholders for your own domain, segment or channel and run it as-is from the console, Slack, or the API.
What comes back
A 200-contact prospect list that matches criteria you explicitly signed off, delivered two ways: a CSV in the conversation where you shaped it, and clean records in HubSpot ready for sequencing. The sample-first checkpoint means the interpretation was corrected when it cost ten rows, and the thread preserves the full decision trail from first description to final import.
Make it yours
- Raise the stakes gradually: approve the first hundred, review reply quality for a week, then reply in the same thread to extend the same criteria to the next batch.
- Run it segment by segment: keep one thread per vertical so each thread's corrections stay clean, and compare sample quality across segments before committing spend to any full build.
- Flip the entry point: start from Intent Data's surging accounts and ask for ten samples of who to contact at each, correcting titles before the enrichment run.
Frequently asked questions
What happens if I never respond to the sample?
Nothing. The full run is gated on an explicit yes in the thread, and silence holds it indefinitely. The mission stays paused with the sample and criteria intact, so you can come back days later, make one correction, and release the run without starting over or re-explaining the target.
Can the corrections get lost between sample and full run?
The thread is the system of record. Every correction you make revises the stored criteria, and the agent restates the final definition before running, so what executes is what you last confirmed. Thread memory persists across the whole exchange, including if it spans multiple days.
What exactly gets written into HubSpot?
Only the approved contacts, as new or updated records with the fields shown in your sample. The agent checks against existing records to avoid duplicates, and it does not modify unrelated properties, lists, or workflows. If your HubSpot has rules you want respected, state them in the thread and they become part of the mission.