How I AIChatGPT agent mode: The “little helper” that transformed recruiting & solved parking nightmares
CHAPTERS
What ChatGPT agent mode is—and why Michal needed it for real work
Claire introduces ChatGPT’s agent mode (agentic browsing) and Michal explains the core problem: talented teams lose hours to repetitive, manual tasks just to gather information. For recruiting, the key requirement was not just web search, but logging into LinkedIn and taking actions like a human would.
- •Agent mode = ChatGPT can browse and perform actions, not only answer in text
- •Recruiting workflow pain: hours spent scanning LinkedIn profiles against criteria
- •Why login matters: internal/behind-login LinkedIn data is required
- •Goal framing: offload “mundane, repeating work” so humans focus on higher-value tasks
Designing the recruiting agent: “little helper” prompt structure and constraints
Michal walks through how he built an effective agent prompt by first setting a role (“You are an IT recruiter”), then defining an explicit task flow, then adding hiring-team-specific constraints. Claire highlights how interviewing colleagues to capture their step-by-step process is often enough to automate meaningful work.
- •Prompt pattern: role → task steps (including ‘let me take over to log in’) → restrictions/criteria
- •Upload the job description as the reference artifact the agent must match against
- •Examples of constraints: Israel-based, LinkedIn-active in last 3 months, seniority/title fit, tenure rules
- •Use of thresholds (e.g., ~70% match) and fixed outputs (e.g., ‘find up to five profiles’)
- •Co-pilot behavior: agent can pause and ask for help when it hits certain steps or issues
Agent mode in action: watching it browse LinkedIn like a human operator
They demo the agent navigating LinkedIn: reading the job description, opening LinkedIn, searching, clicking through results, and evaluating profiles. Claire notes how surprising this “magic computer” experience can be for non-technical users and calls out the UX design choices that make agent browsing understandable and usable.
- •Live browsing: cursor movement, clicking, filtering, opening profiles
- •Transparency: agent narrates its plan/thought process while working
- •Practicality: you can walk away; mobile notifications alert you when it’s done
- •Accessibility: helps people who struggle with complex site UX/filters
- •Product design takeaway: the interface makes agent behavior feel legible, not creepy or confusing
Output format that recruiters can trust: ranked candidates + anonymization
Michal shows an example result: five candidates returned within about ten minutes, with match scores and links. They discuss why scoring helps compare options, and why anonymizing candidate info can reduce bias and focus reviewers on qualifications against stated criteria.
- •Deliverable: table of candidates with profile links and a match score/ranking
- •Match score is not perfect science, but improves triage and comparability
- •Anonymization as a useful recruiting pattern to reduce bias signals (name/school, etc.)
- •Hidden efficiency gain: the agent doesn’t get “distracted” by feeds/notifications like humans do
Team validation and impact: quality wins, not just speed
Michal sanity-checks the AI’s first output by sending it to the hiring manager for evaluation. The manager reports that four of five were strong fits they hadn’t found manually, which builds trust and leads to expanding the workflow across more roles.
- •Human-in-the-loop validation: hiring manager reviews AI-sourced shortlist
- •Outcome: 4/5 candidates were net-new and high quality; 1 already in pipeline
- •Trust-building tactic: start small (5 candidates) before scaling
- •Result: becomes a real part of the hiring process, freeing recruiters for outreach and relationship work
- •Claire’s broader point: AI can improve both speed and quality, especially for “last-mile” discovery
From research PDFs to “talking” personas: the goal for custom GPTs
The conversation shifts from finding real people to creating interactive buyer-persona GPTs. HoneyBook had extensive, expensive customer research trapped in documents, so the goal became making those personas conversational and easy to consult during product and marketing decisions.
- •Problem: persona research exists but is underused because it’s buried in long documents
- •Target outcome: talk *with* a persona, not just ask *about* it
- •Custom GPT basics: name/description matter less than strong instructions + knowledge setup
- •Key realization: instructions must embody identity and behaviors; files alone lead to “about the persona” answers
Using NotebookLM to convert source research into persona prompts
Michal uses Google’s NotebookLM because it can be constrained to user-provided sources and provide citations. He prompts it as an expert prompt engineer to generate highly detailed persona instructions—while explicitly forbidding invention beyond what’s written or implied in the research.
- •Why NotebookLM: answers only from uploaded sources; can toggle sources on/off
- •Citations enable auditing: verify each persona detail traces back to research
- •Prompt approach: role (“expert prompt engineer”) + mission + detailed guidelines
- •Critical constraint: ‘Don’t add or modify text not written or implied’ to reduce hallucinations
- •Workflow: generate persona prompts → save as notes → review/verify with citations
Refining prompts for production: length limits and persona guardrails
The initial persona prompts needed tightening: some exceeded ChatGPT’s custom GPT instruction limits and lacked behavioral guardrails. Michal refines using ChatGPT and sometimes Claude, adding boundaries to keep the personas safe, focused, and consistent at scale.
- •Practical constraint: custom GPT instructions capped (e.g., 8,000 characters)
- •Guardrails added: not a general assistant; avoid follow-ups; keep tone respectful
- •Safety boundaries: avoid political/religious/gender/racial commentary; no distasteful content
- •Reason for guardrails: coworkers will “stress test” personas with off-topic or provocative prompts
- •Technique: use an LLM to compress/strengthen instructions while preserving fidelity to research
Demo: brainstorming with different persona GPTs for marketing copy
Michal demonstrates asking persona GPTs for ad headlines that would catch their attention during a busy workday. Different personas return distinct headline styles and explain the reasoning, letting teams test messaging against customer mindsets quickly.
- •Example question: “What ad headline would catch your attention…?”
- •Personas provide options plus rationale for why each resonates
- •Comparison across personas shows meaningful differences in tone, pain points, and motivations
- •Value: scalable proxy for conversations with thousands of customers, always available
- •Team adoption signal: personas become frequently used internal tools for ideation
Solving SF parking chaos: Giants game schedule → Google Calendar ICS automation
Michal tackles a very practical operations problem: parking near Oracle Park spikes dramatically on day games. He uses ChatGPT to find relevant home games within a time window and generate a filtered ICS calendar file that flags high-risk parking days without blocking availability.
- •Pain point: game-day pricing jumps (e.g., $50/day to $40+/hour) near Oracle Park
- •Need: know day games in advance to choose public transit
- •Prompt asks for: home games only, next 6 months, start time morning–2pm, create ICS
- •Calendar design: all-day events set to ‘Free’ so they don’t mark the user busy
- •Verification step: also output a textual list of included games/dates/times for cross-checking
Lightning round: internal automation roles, prompting habits, and a rewrite template
They close with a discussion of Michal’s role as a Technical Operations Engineer building and enabling internal tooling. Michal shares his go-to prompting technique: ask the model to rewrite the prompt by specifying what’s wrong, what “right” looks like, and explicitly granting permission to delete/replace parts of the original.
- •Role scope: integrate tools, build internal apps/bots, connect systems, enable others via training/docs
- •Perspective: each department operates like a small business that benefits from automation
- •Claire’s take: internal tools teams are newly empowered in the AI era—high impact, faster iteration
- •Prompt-improvement template: provide current prompt + failures + desired properties + permission to rewrite/delete
- •Cultural aside: Michal likes ALL CAPS; practical takeaway is structured iteration over frustration
