How I AI: Automate Recruiting and Build Interactive Personas with Michal Peled of HoneyBook

In this episode of How I AI, HoneyBook technical operations engineer Michal Peled demonstrates a ChatGPT recruiting helper, five research-grounded customer personas, and a shared calendar for avoiding event-day parking prices near the office.

Michal calls these tools "little helpers." Each one starts with an existing manual process, uses the lightest technical approach that fits, and keeps a person responsible for checking the result.

The three workflows use different forms: ChatGPT agent mode for a browser task, NotebookLM plus custom GPTs for customer research, and a generated ICS file for a shared calendar.

source LinkedIn candidates with a recruiting helper

How to Automate LinkedIn Recruiting with a ChatGPT AI Agent: Use a browser agent to produce a LinkedIn sourcing shortlist from a job description and job-related criteria, while keeping login, profile validation, outreach, and every hiring decision with people.

Sourcing candidates is one of those jobs that quietly consumes hours. Recruiters manually search LinkedIn, apply filters, compare profiles against a job description, and repeat the process over and over. Michal saw that friction on her hiring team at HoneyBook and built an AI agent to handle the repetitive parts.

The goal was to take that load off of them. And ChatGPT Agent mode came just in time.

translate the recruiter's process into instructions

The workflow starts with a very structured prompt. Michal gives the AI a role, a detailed task, and clear constraints, essentially mapping the exact workflow a recruiter already follows. To build it, she interviewed her colleagues about how they source candidates step by step, then translated that process directly into instructions.

Here's the prompt she used:

You are an IT recruiter. 
Log into LinkedIn using my account. If not already logged in, let me take control and log in. Find up to five LinkedIn profiles where the current title and job description match the attached job description. 
Restrictions:
- Candidates must be from Israel or currently working at an Israeli company.
- They must be active in LinkedIn within the last three months.
- The current job role must be close enough to the open role in title and seniority.
- The candidates must either work in their current workspace more than a year, or they can be unemployed, but no more than a year, and have worked in their last workplace for over a year.

The prompt includes a handoff for authentication: if LinkedIn is not already logged in, the agent should let Michal take control. The browser agent remains a supervised tool rather than an unattended account operator.

run and supervise the browser agent

ChatGPT agent mode opens a browser environment and shows the actions it is taking as it navigates pages and fills searches.

The progress panel narrates planned and completed actions. It is useful for monitoring the run, but it should not be treated as a complete record of the model's private reasoning.

validate the candidate list with the hiring team

After roughly 10 minutes, the agent returned five candidates with profile links and generated match scores. Michal sent the table to the hiring manager for review.

The feedback was better than expected. Of the five candidates the agent surfaced:

• The manager said four were new prospects who fit the description and were worth approaching.
• The fifth was already in the interview process, which showed that the search could recover one candidate the team had independently selected.

This was one five-candidate test validated by one hiring manager. It justified further trials; it does not establish a general accuracy rate for automated recruiting.

turn customer research into interactive personas

How to Create Interactive AI Personas from Static Customer Research: Convert static customer research into clearly labeled synthetic personas whose responses stay grounded in cited source material, then use them for exploration without treating them as real customer validation.

Most companies invest a lot in customer research, which results in detailed buyer personas. The problem is, this valuable insight often gets trapped in dense PDFs and slide decks that people rarely look at in their day-to-day work. Michal's team at HoneyBook faced this exact issue with five detailed personas they had developed. Her goal was to make them living, breathing entities that the product and marketing teams could actually talk to.

synthesize the source material in NotebookLM

HoneyBook had hundreds of pages of research behind five customer personas. Michal used NotebookLM, a Google tool built on Gemini, to work with the documents for one persona at a time.

She picked Notebook LM for two reasons:

1. Source selection: NotebookLM can restrict an answer to the sources selected in the notebook instead of retrieving unrelated web material.
2. Citations: Responses point back to passages in the uploaded sources, giving Michal a way to inspect the basis for a claim. Citations support verification; they do not guarantee that every synthesis is correct.

ask NotebookLM to draft persona instructions

Michal asked NotebookLM to act as a prompt engineer and draft detailed custom-GPT instructions from the supplied research. She did not simply upload every document to one general chatbot.

Here's the core of her prompt to NotebookLM:

You are an expert prompt engineer specializing in creating custom GPTs by providing strong AI prompts. Your mission is to create AI prompts for custom GPTs representing entrepreneurs and small business owners... you'll craft highly detailed nuance and authentic ChatGPT prompts for five distinct buyer personas based on your sources.
Guidelines:
- Ensure that the prompt correctly and fully describe the core identity, mindset, decision making style, tone and communication style...
- ...business needs and the technology stack and the journey maps, social media preferences...
- Don't add or modify text that is not written or implied in the text. I know you're creative. I am turning you down. The text describe a specific persona must remain true to the original persona.

I laughed at that last instruction. The "don't make up stuff" directive belongs in basically everyone's prompt toolkit.

refine the prompts and add boundaries

Michal reviewed the drafts, used ChatGPT and Claude to tighten them, kept each under the 8,000-character custom-GPT instruction limit, and added explicit behavioral boundaries.

She added instructions like this to keep the personas professional and tightly scoped:

You do not act as a general purpose assistant. You do not ask follow up questions, you avoid slang, bad language or distasteful content and keep communication respectful and inspiring. You avoid political, religious, gender, or racial commentary.

This is a smart pattern for enterprise GPTs in general. Users will absolutely try to push personas off the rails, intentionally or not. These guardrails keep the interaction useful instead of chaotic.

make each persona available as a custom GPT

With the final prompts ready, Michal created five custom GPTs, one for each persona. Instead of digging through research PDFs, her coworkers can now directly ask a persona like "Balanced Blake" questions in plain language.

For example, when asked "What kind of ad headline would catch your attention during a busy workday?", Blake responds:

A few would catch my eye. 'Save 10 hours a week with this tool. No tech skills needed' or 'From chaos to clarity. One dashboard to run it all.'

Different persona GPTs return different messaging ideas based on their underlying research. They are tools for exploring documented customer perspectives, not substitutes for new customer interviews or validation.

create a calendar for daytime Giants games

How to Create a Custom Google Calendar from a ChatGPT Prompt: Create a purpose-built shared calendar from an authoritative event schedule by filtering the exact dates that matter, generating an ICS file, and validating every event before import.

This last workflow is a personal favorite because it solves such a specific, tangible, and frustrating problem. The HoneyBook office is located right next to Oracle Park, home of the San Francisco Giants. On game days, parking rates skyrocket from a flat ~$50/day to a whopping $40+ per hour.

As someone whose former office was also behind the ballpark, I deeply relate to this problem. I once paid nearly $100 to park for a meeting and immediately texted a friend in disbelief. Michal's team kept getting blindsided by game-day pricing, so she built a simple AI workaround.

The one-shot request

This one didn't require a complicated agent workflow. Michal just needed a reliable way to flag weekday daytime games ahead of time. She went to ChatGPT with a very straightforward prompt:

Find all home games that take place in Oracle Park in San Francisco during the next six months. Filter out only the games that start anywhere between morning to 2:00 PM. Using these dates create an ICS file for Google Calendar that will show these dates as an all day event. Availability: free. The event description should contain the game details and time. Also provide a textual list of all the dates, times, and events included.

Michal asked for an ICS file containing only games that began between morning and 2 p.m., displayed as all-day events with availability set to free. She also requested the game details and a separate text list so she could verify the generated calendar.

The generated calendar

ChatGPT returned an ICS file and a text list after 36 seconds in the recorded run. Michal imported the file and shared the calendar with teammates. The text list gave her a second surface for checking dates and times before relying on it.

The workflow is deliberately small: it converts a public schedule into a team-specific warning without blocking anyone's work calendar.

The operating pattern

The recruiting helper, persona tools, and parking calendar address three different kinds of operational friction. In each case, Michal starts with the actual process and adds a human checkpoint where the output could affect a candidate, customer decision, or colleague's plans.

The common test is whether a small tool removes repeated work while leaving its evidence and review path visible. That is a stronger standard than whether the demo looks impressive.