How I AIChatGPT agent mode: The “little helper” that transformed recruiting & solved parking nightmares
CHAPTERS
- 0:00 – 2:18
Agent mode in ChatGPT: what it is and why it’s useful for real work
Claire introduces ChatGPT “agent mode” as a step beyond chat: it can browse and take actions on websites. Michal frames the core value as offloading repetitive, manual work so teams can spend time on higher-value decisions.
- •Agent mode can browse the web and perform step-based actions, not just answer in text
- •The motivation: reduce mundane, repetitive tasks that drain expert time
- •Framing an agent as a “little helper” sets the mental model for better prompts
- 2:18 – 5:16
Recruiting pain point: speeding up LinkedIn sourcing without losing specificity
Michal explains the hiring-team workflow of scanning many LinkedIn profiles against a job description and why it’s time-consuming. Agent mode is a fit because it can log in, run searches, open profiles, and apply team-specific criteria.
- •Recruiters spend hours manually filtering candidates on LinkedIn
- •Agent mode can work inside authenticated sites (e.g., LinkedIn)
- •Goal: search profiles and evaluate against hiring-team requirements
- •Automation supports recruiter focus on outreach and relationship-building
- 5:16 – 10:58
Designing the recruiting agent prompt: role, task steps, and hiring restrictions
They break down the prompt structure that makes the agent effective: set a role (“IT recruiter”), define the task clearly, and then encode the team’s real screening criteria. The episode emphasizes interviewing colleagues to capture their true step-by-step process.
- •Start with role assignment (e.g., “You are an IT recruiter”)
- •Explicit workflow steps: log in, search, find up to five matching profiles
- •Encode real screening criteria (location, activity recency, tenure/unemployment rules)
- •Specify outputs: number of candidates and match threshold/score
- •Human-in-the-loop control points (agent pauses for login or issues)
- 10:58 – 16:36
Live demo: watching the agent browse LinkedIn and narrate its reasoning
Michal shows the agent reading the job description, navigating LinkedIn, and clicking through results while exposing its internal ‘thoughts.’ Claire highlights how this interaction design expands accessibility and lets users delegate tedious navigation.
- •Agent reads the provided job description then opens LinkedIn
- •On-screen cursor movement + narrated reasoning makes behavior inspectable
- •You can let it run while you do other work; phone notifications when done
- •Broader implication: agents make complex UIs usable for less technical users
- 16:36 – 19:55
Evaluating outputs: match scores, anonymization, and staying focused vs. getting distracted
They review a sample output table of five candidates with match scores and discuss why ranking helps compare results. Claire notes the value of anonymizing candidates to reduce bias, and jokes that agents don’t get distracted by feeds like humans do.
- •Output includes five candidates plus match scores for comparability
- •Match scores aren’t perfect science but help triage effort
- •Anonymized profiles can reduce bias in early screening
- •Agents are efficient because they don’t get pulled into notifications and feeds
- 19:55 – 23:50
Team reception and impact: faster sourcing with surprisingly strong candidate quality
Michal validates early results with a hiring manager, who finds most candidates were new and strong fits. This shifts the agent from experiment to an expected part of the recruiting process, freeing recruiters for higher-value work.
- •Pilot validation: hiring manager reviewed the five profiles for fit
- •4/5 candidates were new finds; 1 had already been discovered manually
- •Success changed internal appetite: run it across more roles and expand output size
- •Counterpoint to skepticism: speed gains can also improve quality
- 23:50 – 28:51
From ‘finding people’ to ‘creating people’: turning persona research into interactive GPTs
The episode pivots to customer research: five buyer personas existed in long documents that teams rarely used. Michal’s goal is to transform static research into interactive persona GPTs that teams can ‘talk to’ for product and marketing decisions.
- •Problem: expensive research trapped in PDFs/docs, rarely referenced
- •Goal: create five custom GPTs that behave like the personas themselves
- •Key distinction: talk with a persona vs. asking about a persona
- •Instructions (not just uploaded files) must embody beliefs, behaviors, and preferences
- 28:51 – 34:52
Using NotebookLM to extract persona prompts grounded only in source documents
Michal chooses NotebookLM because it answers strictly from uploaded sources and provides citations. He prompts it as an expert prompt engineer to generate detailed persona instructions without inventing facts or filling gaps from general knowledge.
- •NotebookLM can constrain answers to provided sources (reduces hallucinations)
- •Ability to toggle sources on/off to control what informs answers
- •Citations enable verification of persona details back to original research
- •Prompt includes explicit persona dimensions: mindset, decision style, tone, tech stack, journey, social prefs
- 34:52 – 37:05
Guardrails and refinement: tightening prompts for limits, safety, and realism
The initial persona prompts needed refinement: they were too long for Custom GPT instruction limits and lacked guardrails. Michal uses ChatGPT/Claude to compress and add behavioral constraints (e.g., no politics, no profanity, no follow-up questions).
- •Custom GPT instruction limit (~8,000 characters) requires concise prompts
- •Add guardrails to prevent off-topic or adversarial employee testing
- •Behavior constraints: respectful tone, avoid sensitive topics, avoid slang/distasteful content
- •Reinforces “don’t make things up” principle when grounding in research
- 37:05 – 41:19
Persona GPT demo: testing ad headlines across different customer archetypes
Michal demonstrates asking a persona what ad headline would catch attention, then compares outputs across different personas. The value is rapid, repeatable “customer perspective” checks—available anytime—based on consolidated research.
- •Ask personas questions like: headline ideas, onboarding impressions, churn prevention
- •Different personas produce distinct, consistent answers aligned to their archetype
- •Personas become a shared internal brainstorming tool used frequently by colleagues
- •Transforms thousands of customer insights into an interactive decision aid
- 41:19 – 48:49
Parking price nightmares at Oracle Park: turning game schedules into a filtered calendar
Michal addresses a San Francisco-specific pain: parking spikes on Giants game days near the office. He uses ChatGPT to find upcoming daytime home games and generate a shareable ICS calendar so employees know when to take public transit.
- •Problem: parking jumps from day rate to $40+/hour on daytime game days
- •Need: identify home games at Oracle Park with start times before ~2 PM
- •ChatGPT generates an ICS file + a textual list for verification
- •Calendar events set to ‘Free’ availability so they don’t block schedules
- 48:49 – 58:45
Lightning round: the internal tools role, prompt debugging tactics, and closing advice
They close with a discussion of Michal’s Technical Operations Engineer role and why internal tools teams can move fast with AI. Michal shares a practical technique: ask the model to rewrite your prompt using explicit failure descriptions, success criteria, and permission to delete/replace anything.
- •Internal tooling blends no-code, automation, integrations, and AI apps (e.g., Slack bots)
- •Emphasis on enablement: training, docs, advisory, helping teams self-serve
- •Prompt improvement template: include what’s wrong, what ‘right’ looks like, allow major rewrites
- •Connect on LinkedIn; use agents to even improve your own LinkedIn profile fit
