# 21 Applications on the First Run: How Devin Huber screens podcast guest applications with a monthly agent that can only act through n8n

Devin Huber, an AI automation engineer in Columbus, Ohio, built a guest application pipeline for the podcast of the leadership-development institute where he works. Once a month a scheduled Claude task scores new guest applications against a 15-point rubric, checks whether the CEO already knows the applicant, writes a personal decline for weak fits and sends the CEO a digest. The model never touches the sheet or the mailboxes directly: every read and send goes through a no-AI n8n workflow with named actions. Its first run cleared a backlog of 21 applications, and it has processed 37 in all.

**Try this step:** <strong>Daily intake, no AI.</strong> An n8n workflow reads new applications, skips any it has already handled using 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.


| | |
|---|---|
| **Who** | Devin Huber, AI automation engineer (Columbus, OH), who documents his builds at devinhuber.com. Name only (no social profile verified) |
| **Problem** | Every guest application is a judgment call, strong fits were getting buried among weak ones, and every decline deserved a personal reply that took time to write |
| **Stack** | A scheduled Claude task for the monthly batch, a no-AI n8n workflow with named actions for every sheet and mail step, Google Sheets, n8n data tables, an exclusion list and one emergency stop |
| **Confidence** | Medium (first-person, from the agent's own sheet; hours saved are projected) |
| **Last verified** | 2026-11-13 |

## The problem
Screened by hand, strong podcast guests got lost in the pile, and a weak fit still deserved a reply that wasn't a form letter. A wrong decline carries its own risk: turning down someone the CEO already knows costs more than replying late. He first built the screener as two Zapier agents, split to stay inside Zapier's execution limits, with a written build sheet and test plan, then rebuilt it as one agent in a scheduled Claude task with every sheet and mail step in n8n.

## The blueprint (as he describes it)
1. **Daily intake, no AI.** An n8n workflow reads new applications, skips any it has already handled using 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.
2. **A monthly batch.** A scheduled Claude task reads the month's applications and the submissions log through the n8n workflow.
3. **Score against a rubric.** Five criteria, scored out of 15, plus the agent's own judgment.
4. **Check 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. A name only counts when the full name appears, because the mail search is not a phrase search, and the employer's own automated replies are ignored.
5. **Decline or hold.** Weak fits get a personal decline written for them, and the n8n action that sends it refuses while the emergency stop is on. Strong fits and anyone held go to the CEO.
6. **One digest.** The CEO gets one monthly email summarizing the pipeline, and its recipient is fixed in the workflow.

## Stated results
- A backlog of 21 applications cleared on the first run, and 37 processed in all.
- He counts that as 37 hours saved so far and about 100 hours a year at the current rate, using 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, less the CEO's 45-minute monthly review of the digest.
- The agent scores every application and prepares a personal decline for weak fits, and the CEO decides anything held.
- It is running in production at his current employer.

## Caveats
- The 100 hours is a projection, "not a timing," based on 6 applications in September and 10 in October.
- How often leadership overrules the agent's scores is not measured yet.
- The n8n workflow does not itself refuse a second decline to the same person; that rests on the task checking the record first.
- Application text is treated as untrusted input, with defenses against instructions hidden inside it, but the page does not say whether any attack has been tried.
- It is a portfolio write-up: he says he is open to AI automation roles, client names are withheld, and the employer and podcast are not named.
- Handles: name only. No social profile is linked from his site, so none is tagged.

## FAQ
### How does Devin Huber screen podcast guest applications?
A scheduled Claude task scores the month's applications on a 15-point rubric, checks for existing relationships with the CEO, writes personal declines for weak fits and emails the CEO a digest. Every sheet and mail step runs through a no-AI n8n workflow.

### What happens if the relationship check fails?
The applicant is held for the CEO, not declined. He designed it to fail closed because a wrong decline to someone the CEO knows costs more than a late one.

### How much time does it save?
He counts 37 hours so far and projects about 100 hours a year from the team's own time per application. He labels it a projection, not a timing.

### Can the model send email on its own?
Only through named n8n actions with fixed recipients, and the decline action refuses while the emergency stop is on.

## Sources
- https://devinhuber.com/projects/guest-flywheel/ (primary: Devin Huber's first-person project page, "Guest Application Flywheel", checked Nov 13, 2026)
- https://devinhuber.com/ (his site: AI automation engineer, Columbus, OH)
- https://devinhuber.com/builds/ (his list of builds at his employer)

---
*Want agents like these working your distribution? [Get It Built](/get-it-built). Get new teardowns weekly: [subscribe](/subscribe) (double opt-in).*