Guest Application Flywheel
Once a month a Claude Cowork task scores new podcast guest applications against a rubric, holds anyone the CEO already knows, sends weak fits a personal decline on its own and emails the CEO a digest. Every sheet and mail step goes through a no-AI n8n workflow with named actions.
For a businessCuts hours of manual screening so strong podcast guests aren't lost in the pile, while a person keeps the final say on every decision.
- Status
- In production at my employer
- Result
- About 110 hours a year saved at 10 applications a month, by the team’s own time per application: a projection, not a timing
- How it was measured
- The agent’s own sheet counts the time per application it screened; the yearly figure applies that to the expected volume
- Biggest limit
- How often leadership overrules its scores isn’t measured yet
01 · Context The problem
Every guest application is a judgment call: is this person a fit for the show, and if not, how do you say so without a form letter? Screened by hand, strong fits got buried among weak ones, and every decline deserved a personal reply that took time to write.
02 · Pipeline How it works
- Intake, daily, no AI
An n8n workflow reads new applications, skips any it has already handled (a ledger keyed on time, name and email) and records each one once. While the emergency stop is on it does nothing, and the rows wait for the next run.
- Monthly batch
A Claude Cowork scheduled task reads the month’s applications and the submissions log through the n8n workflow.
- Score against the rubric
Five criteria, scored out of 15, plus the agent’s own judgment.
- Check for existing relationships
Before any decline, code searches the CEO’s mail, the shared inbox and her contacts for the applicant’s email, company domain and full name, and checks an exclusion list. Any real hit, or a search that fails, holds the applicant for the CEO.
- Decline or hold
Weak fits get a personal decline written for them and sent automatically; the n8n action that sends it refuses while the emergency stop is on. Strong fits and anyone held go to the CEO.
- Monthly digest
The CEO gets one email summarizing the pipeline; the recipient is fixed in the workflow.
03 · Choices Design decisions
Claude decides, n8n acts
The model reaches the sheet and the mailboxes only through one unpublished n8n workflow with seven named actions. The intake tab is read-only to it, and the digest recipient is fixed.
Fail closed on relationships
If the relationship search errors, the applicant is held, not declined. A wrong decline to someone the CEO knows costs more than a late one.
Applications are data
Application text is treated as untrusted input, with defenses against instructions hidden inside it.
One emergency stop
A single setting pauses the outbound steps: the decline action and the daily intake both check it before doing anything.
04 · Evidence How I tested it
Built first as two Zapier agents, split to stay inside Zapier’s execution limits, with a written build sheet and test plan. Rebuilt as one agent in a Claude Cowork task, with every sheet and mail step in no-AI n8n.
The relationship check counts a name only when the full name actually appears, because the mail search is not a phrase search, and it ignores the employer’s own automated replies.
05 · Outcomes What it has done so far
- Cleared a backlog of 21 applications on its first run; 27 screened in all. At about 10 applications a month, that is about 110 hours a year saved, by the team’s own time per application: 20 minutes to read one, 10 to research it and an hour more when it needs deeper research (14 of the first 27 did), less the CEO’s 45-minute monthly review of the digest. A projection from estimates, not a timing.
- Leadership no longer handles applicant email: the agent scores every application and declines weak fits on its own, and the CEO decides anything held.
- Running at my current employer without interruption.
06 · Security Untrusted input and what the model can touch
- Application text is untrusted input, with defenses against instructions hidden inside it.
- The model can call only seven named actions; recipients and the intake sheet are fixed in the workflow.
- Sending is refused while the emergency stop is on, and a failed relationship search holds the applicant.
07 · Honest notes Known limits
- The n8n workflow does not itself refuse a second decline to the same person; that rests on the task checking the record first.
- The ~110 hours is projected from the team’s own time per application at about 10 a month (October brought 12), not timed.
Want to see this one run?
I'll screen-share the workflow, the test set and the results, including what didn't work.