ChatGPT agent mode: The “little helper” that transformed recruiting, crafted user personas, and solved parking nightmares | Michal Peled (Honeybook)

Michal demonstrates using ChatGPT's agent mode to automate LinkedIn recruiting by having the AI log in, search profiles, and filter candidates based on specific criteria. The workflow reduced hours of manual browsing to 10 minutes and found 4 high-quality candidates that manual searches missed. Key insights include structuring prompts with role definition, task description, and specific restrictions that mirror actual recruiter workflows.

• •Define the agent's role first ("you are an IT recruiter") before describing the task to create a focused helper
• •Add specific restrictions from actual hiring team workflows: location requirements, LinkedIn activity within 3 months, tenure requirements (1+ years in current role or unemployed less than 1 year)
• •Include verification steps like match scores (70%+ threshold) and ask for both the output file AND a textual list to validate results
• •Agent mode allows the AI to perform actions (login, click, navigate) while showing its reasoning in real-time
• •Out of 5 candidates found in 10 minutes, 4 were never discovered manually and 1 was already in the interview pipeline - proving quality over speed

Michal created five interactive AI personas from hundreds of pages of customer research using NotebookLM and custom GPTs. Instead of uploading files to ChatGPT, she used NotebookLM to generate tight instruction prompts that embody each persona's identity, allowing teams to have real conversations with representative customers 24/7. The personas are now among the most-used internal tools at HoneyBook.

• •Use NotebookLM to transform research documents into persona prompts because it only answers from provided sources and includes citations for verification
• •Prompt NotebookLM as an "expert prompt engineer" with specific guidelines: core identity, mindset, decision-making style, business needs, tech stack, journey map, social media preferences
• •Critical instruction: "Don't add or modify text that is not written or implied in the source" to prevent hallucinations
• •Instructions should make the GPT BE the persona, not talk ABOUT the persona - use "you are" language throughout
• •Add guardrails to custom GPTs: avoid political/religious/gender/racial commentary, no slang or bad language, don't act as general assistant, don't ask follow-up questions

Demonstration of how different personas respond uniquely to the same questions about ad headlines and product features. Each persona provides answers based on their specific characteristics from the research, representing thousands of potential customers. Teams can now test campaigns, features, and messaging against personas instantly instead of waiting for expensive research cycles.

• •Same question ("what ad headline would catch your attention") yields completely different answers from different personas based on their pain points and preferences
• •Personas explain WHY certain headlines work for them, providing actionable insights beyond just preferences
• •Use cases include testing ad campaigns, product onboarding flows, feature prioritization, and churn prevention strategies
• •Personas are available 24/7 for quick validation without scheduling research sessions or waiting for survey results
• •The personas have become highly-used internal tools, showing teams actually engage with research when it's interactive

Michal solved a San Francisco-specific problem where parking near HoneyBook's office spikes from $50/day to $40+/hour during Giants games. She used ChatGPT to create a custom ICS calendar file that filters only morning/afternoon home games and marks them as all-day events with availability set to free, so the team knows when to take public transit instead of driving.

• •Prompt ChatGPT to find specific events (Oracle Park home games), filter by time (morning to 2PM only), and output in specific format (ICS file for Google Calendar)
• •Set events as "availability: free" so they don't block your calendar as busy - just provide ambient awareness
• •Always request both the output file AND a textual verification list to validate the AI's work before importing
• •Include event descriptions with game details and times for manual verification of accuracy
• •Simple prompts can solve hyper-specific local problems that no existing product addresses - don't need complex role definitions for straightforward tasks

Michal describes her role as building internal tools, automations, and integrations while also teaching and enabling others. She views each team at HoneyBook as a small business with their own goals and expertise, and her job is to remove friction so they can focus on what they love. This represents the evolution of internal tools teams in the AI era.

• •Technical operations engineers research/integrate paid tools, build custom internal tools using no-code and coded solutions, and create integrations between systems
• •Focus on enabling others through company-wide presentations, personal advisory, training classes, and documentation - not just building for them
• •Internal tools teams are having their moment with AI - can move faster than customer-facing products due to fewer legal/compliance constraints
• •Treat each department as a small business with unique needs, goals, and expertise that deserves custom solutions
• •The goal is eliminating manual, thoughtless, repetitive tasks so teams spend more time on high-value work they're passionate about

When AI isn't responding correctly, Michal's go-to technique is using ChatGPT to improve the prompt itself. She provides the current prompt, outlines what's wrong with the output, specifies how she wants it to be right, and explicitly gives permission to delete, rewrite, or add anything. This meta-prompting approach usually fixes issues in one iteration.

• •Use all caps for emphasis (despite debates about whether AI cares) - psychological benefit for the prompter matters
• •Structure prompt improvement requests: current prompt + what's wrong (inconsistent, hallucinations, etc.) + how you want it to be right
• •Explicitly give permission to "delete, rewrite, or add anything" - ChatGPT tends to be a pleaser and won't change your work without permission
• •This technique works for fixing your own prompts and helping colleagues debug their custom GPTs
• •Usually only takes one iteration to get a dramatically better prompt that works as intended