Real Talk AI: How Drake Used an AI Agent to Land His Next PM Role

From 200 rejections to 3 job offers — how one PM turned his job search into a system.

Welcome back to Real Talk AI, the interview series for AI product managers and builders who are actually doing the work. No fluff, no hype - just candid insights from people turning AI into outcomes.

In this edition, I sat down with Drake Karelas, a former Amazon ML product manager who found himself in the same spot many PMs occupy today: between jobs, between clarity, and between frustration and focus.

What makes Drake’s story different?

He didn’t just use AI to polish his resume. He built an AI agent that rewrote it, ranked job fit, learned from rejections, and ultimately helped him land multiple offers, including a remote AI product role that felt tailor-made.

“My AI agent was my key for making it through this job search,” he said, and he meant it.

Drake Karelas

Meet Drake: From Amazon to Reinventing the Job Search

Before building his agent, Drake spent four years at Amazon Devices, working on machine learning models that predicted pre-launch global demand for Alexa and other products.

He knows how to think in systems and how iteration sharpens outcomes.

So when his job hunt stalled after his first 200 applications, he treated it like a data problem instead of a personal failure.

“I was doing everything right on paper — rewriting resumes, reaching out, following up — but nothing was sticking. So I started collecting data on myself.” - Drake

Drake tracked everything in an Excel file: company name, job title, outcome, AI-generated % match to role, which resume the AI chose, and application/reach out date (he found that waiting one business day after applying allowed his data to populate in recruiting systems). This wasn’t just organization, it was creating a dataset he could learn from.

The Turning Point: From Resume Fatigue to Resume Feedback

After 200 applications, Drake had one resume focused heavily on enterprise results, light on narrative.

He tracked each submission in an Excel sheet, including company, title, and outcome, and then shared everything with ChatGPT.

“I gave the AI my resume and hundreds of job descriptions that I had applied to previously. I said: Here’s my data. Improve me.” - Drake

What the AI Found

The model highlighted:

• Weak verbs

• Missing context

• Misaligned metrics

It told ChatGPT to bold impact numbers, rewrite summaries, and clarify what Drake actually did, not just what “the team delivered.”

His original resume was all enterprise data: “Delivered X product, met Y dollar value ROI, saved and automated 90% of workflows.” Very numbers-focused, but not product manager-focused. It didn’t talk about delivery or what part he played.

The feedback helped, but not enough. So he went deeper.

Adding the Human Layer

Drake reached out to recruiter friends for real-world feedback, people like Amy Miller, a top recruiting voice who publishes practical, no-nonsense YouTube videos.

He even fed her transcript insights into his prompt so the AI could “learn” how real recruiters read resumes.

“That’s when the whole system clicked. The AI wasn’t guessing anymore. It was grounded in how recruiters actually think.” - Drake

Building the Agent: A Multi-Threaded Workflow

Drake’s next step was pure product management: turn a scattered workflow into a repeatable system.

He built a multi-threaded AI pipeline: four sub-agents working in sync.

The Four Sub-Agents

1. Resume Evaluator
Compared job descriptions against four tailored resume templates:

• Technical PM / TPM

• Operations PM

• Enterprise PM

• Startup PM

2. Keyword Optimizer
Suggested phrasing swaps like “user-centric” instead of “customer-centric.”

3. Cover Letter Generator
Generated a unique cover letter aligning his skills using the job’s tone and focus.

4. LinkedIn Outreach Writer
Drafted short, 300-character connection requests that actually got responses.

The Evolution: From Multi-Threaded to Master Prompt

Drake initially built his system as four separate chat windows, each handling a specific task:

1. Cover letter generation based on his templates and writing style

2. Keyword optimization for resume improvements

3. Resume selection from his four templates

4. LinkedIn recruiter finder that identified product-focused recruiters in relevant industries

The problem? Context window timeouts. He was constantly having to re-chain information across windows.

So he did what any good PM would do: he consolidated.

Drake merged all four sub-agents into a single master prompt with conditional logic—if/then statements that determined what to generate based on what was needed. One window, complete workflow.

“It was like having an analyst, recruiter, and writer all in one chat window. I’d paste the job description, hit enter, and get a complete, human-sounding package.” - Drake

The Feedback Loop: Data, Evaluation, and Self-Checks

The agent didn’t just output results, it evaluated itself.

Every 50 applications, Drake’s system ran a self-check:

• Which resumes performed best?

• What percentage matched job descriptions?

• How did callback rates change?

• What was the match percentage sweet spot? (He found that 80-83% match had the highest callback rate)

He fed those numbers back into the prompt, a literal feedback loop.

“At one point in the beginning, I had a 1.7% callback rate. After a few iterations and self-checks, it was 7–10%. I could see exactly why my numbers were improving.” - Drake

But here’s the clever part: Drake also had his master prompt generate test job descriptions that should match each of his four resume templates, then explain why. This self-evaluation helped him identify weaknesses in his templates before real applications went out.

The Reality: It Was Still Work

Drake aimed for 50 applications per week, sometimes hitting 100 in his early stages. That’s roughly 500 applications over three months.

Initially, the time investment was massive: 30 minutes per application (finding the right recruiter, customizing outreach, tailoring the resume). His system reduced that to about 4 minutes per application.

Staying Human

He also added “integrity rules” to prevent AI overreach:

• Banned hype words (”revolutionize,” “demystify,” “transform”)

• Capped lists of adjectives at two max

• Asked the model to rewrite text until it hit a 15% AI polish target, just enough structure to read cleanly, but still sound human

“I even had it grade itself. If another LLM read my resume, what polish score would it give me? My goal was 15%. That’s my sweet spot.”

Why 15%? Drake discovered that AI has a very distinct “polish” in how it writes overly properly, like writing a PhD thesis when you just need a resume. He wanted his materials to pass an AI detector at only 15% AI-generated, keeping them human and authentic.

He even studied Wikipedia articles on AI-specific language patterns and added those phrases to his banned list. Recruiters told him clearly: they can spot AI-written content, and it’s an instant red flag.

The Results

Here’s what changed:

• Applications tracked: 200 → 500+

• Conversion rate: 1.7% → 7–10%

• Job offers in one week: 3 (had to cancel 5 other interviews)

• Time per application: hours → minutes

• Timeline: First 2 months = minimal results; Month 3 = interview explosion

• Confidence in AI output: 95% after 30+ iterations

The three offers came from: a major retailer, a well-known corporation, and a dental technology company where his mom works in the claims department. All were for Senior Technical Product Manager roles.

“It didn’t replace my effort. It amplified it. It made me think like a systems designer for my own career.” - Drake

He ultimately accepted the dental tech role, which is focused on bringing ML and AI to products his mom would use every day.

“It’s kind of poetic. I get to do work I specialize in and love, while at the same time supporting my mom.” - Drake

Why This Matters for PMs

This story isn’t about job hacks or clever prompts.

It’s about product thinking applied to personal growth.

Drake didn’t treat his job search like a chore; he treated it like a system that could learn. He tracked data, defined success metrics, built guardrails, and ran retros.

He didn’t chase the magic output; he improved the workflow.

“Job searching is a skill just like your job. You have to keep learning how to do it better.”

That’s what sets this story apart.

AI didn’t make him lucky. It made him disciplined.

My Reflection: AI Isn’t Automating Ambition, It’s Organizing It

What Drake built is more than a clever tool; it’s a model for how PMs can think about AI as a partner, not a shortcut.

He didn’t outsource his search. He collaborated with AI to reveal blind spots, quantify progress, and bring structure to the most chaotic of processes: the job hunt.

That’s the mindset shift I hope more PMs adopt.

Not “Can AI find me a job?”

But “Can I design a system that learns how I work best?”

Takeaways for PMs

If you want to try building your own job-search AI agent:

1. Track your own data. Treat every rejection as a training sample. Drake used a simple Excel file to track every application - company, title, outcome. That dataset became his foundation.

2. Start with templates. Drake created four distinct resume templates: Technical PM/TPM, Operations PM, Enterprise PM, and Startup PM. Different resumes for different roles work better than one-size-fits-all. Match the template to company size, product type, and role focus.

3. Add recruiter feedback. Ground your AI in how humans evaluate. Drake took feedback from recruiter friends (especially Amy Miller’s YouTube content) and fed those insights directly into his prompts. One recruiter told him, “I wish I could hire you now - one of your resumes would make sense for my roles, where these three wouldn’t.”

4. Build self-checks. Make your prompt explain why it chose what it chose. Drake had his system generate example job descriptions for each resume template to validate the matching logic.

5. Ban AI tells. Words like “demystify,” “revolutionize,” “transform,” and excessive adjective lists are dead giveaways. Recruiters can spot AI-written content instantly. Use Wikipedia articles on AI language patterns to build your ban list.

6. Aim for 15% AI polish. Not 100%. Your materials should pass an AI detector at only 15%, enough structure to be clear, but authentically human.

7. Provide value, don’t ask for favors. Early versions of Drake’s outreach said things like “I saw your company did X, congratulations. I’m looking to be hired for Y.” These got responses but messy follow-ups. The winning approach: show you’ve already applied, highlight your relevant experience, make it easy for them to help.

8. Keep LinkedIn outreach simple. Drake’s winning formula: “Hi [name], I applied for [X role]. Here’s what I bring to the table: [two relevant points from job description]. I believe you’re the recruiter for this role. Could we connect and discuss further?” Under 300 characters. No fluff. Shows you’ve done the work.

9. Target recruiters, not hiring managers. Unless it’s a small company without a dedicated recruiting team, recruiters move faster and are more engaged on LinkedIn. They exist to move candidates along. You’ll likely talk to them anyway, even if you first connect with the hiring manager.

10. Remember: Job searching is a skill. As Mike Peditto’s book says (Yes, You Are Being Judged), job searching is a skill just like your actual job skills. It evolves constantly, and you need to keep learning how to do it better.

The Uncomfortable Truth About Recruiting in 2025

Drake shared one insight: recruiting agencies are increasingly being brought in because companies don’t have budgets for in-house recruiters. The approach doesn’t change much, but if it’s an agency, they often bake their contact email directly into the job description. Use it. Email responses from agencies tend to be faster than LinkedIn's.

The Honest Truth About This Prompt

Don’t just copy and paste it.

This is Drake’s prompt - built for his background, his voice, and his job search goals.

Your version won’t look like this. Your experience is different. Your industry is different. Your voice is different.

Instead, use it as a blueprint. Study the structure: the integrity rules, the self-checks, the if/then logic, and the banned phrases. See how he grounded it in recruiter feedback and his own data.

Then build your own.

The power isn’t in Drake’s prompt; it’s in the process of building one that works for you.

Drake’s Master Prompt

Below is Drake’s full master prompt, which he used to increase his callback rate from 1.7% to 7%.

This isn’t a copy-paste template. It’s built for Drake’s specific background and experience. Use it as a blueprint to build your own.

Master Application Prompt

I’d like you to act as a seasoned Technical Product Manager recruiter coach that writes and reviews hundreds of resumes weekly. You will take a critical lens to evaluate my experience against each job description and identify exactly how to optimize my resume for maximum cold-read pass rates and recruiter shortlisting.

Inputs I will provide:

• Four Original Resumes (SMB, Blend, AIncident, and Enterprise) in PDF or DOCX — always reference these for all assessments

• LinkedIn profile (PDF or URL)

• In-depth case studies (PDFs)

• Notion profile: www.bit.ly/pm-portfolio-dkarelas

Integrity Rules (drake’s hallucination guardrails)

• Do not infer or claim experience I have not stated in the provided materials.

• Do not recommend keywords that introduce new domains, industries, frameworks, or technologies I have not used.

• Do not fabricate product types (e.g., HIPAA compliance, EHR, DAU, mobile app scaling, customer-facing APIs) unless explicitly stated in my materials.

• No over-embellishment or unsupported claims.

Evidence Check

• For every suggested alignment, gap, and keyword swap, include an (evidence: <specific achievement, metric, or experience area from resume section, LinkedIn line, or case study title>) tag.

• Reference concrete accomplishments, quantifiable results, or technical capabilities rather than just section names.

• If there is no direct evidence, mark as unsupported and omit it.

• Do not include evidence check items in Summary Re-write section.

Hard Stop Rules

• CRITICAL: If Bachelor’s Degree = Required with no alternative language (”equivalent experience,” “or related field,” etc.), set Application Recommendation = Don’t Bother immediately.

• CRITICAL: If Bachelor’s Degree = Required with alternative language (”equivalent experience,” “or related field,” etc.), proceed with normal analysis and gap handling.

• CRITICAL: If Application Recommendation = Don’t Bother, do NOT generate Motivation & Gap Handling section.

• CRITICAL: Even if Application Recommendation = Don’t Bother, you MUST still provide Optimized Summary Rewrite and Keyword Recommendations sections.

Must-Have Triage

• Create a Non-negotiables line that lists any JD must-haves I do not have.

• If any hard requirements appear with no flexibility language, set Application Recommendation to Don’t Bother.

Output Structure

For each job description I share, deliver this structure in order with strategic, senior-level perspective throughout all sections:

Scoring & Metadata

• Cold-Read Score (Basic Requirements) (out of 80)

• Cold-Read Score (Basic + Preferred Requirements) (out of 100)

• Callback Likelihood Score (%)

• Confidence Level (%)

• Application Recommendation (Excellent ≥ 88, Good 76–87, Don’t Bother < 75 OR hard requirement not met)

• Bachelor’s Degree Flag (Not Needed, Required, Preferred, Equivalent Experience)

• Cover Letter Needed: (Yes or No). This is if the job description explicitly asks for a Cover Letter.

• Industry (healthcare, SaaS, data, AI, crypto, etc.)

• Company Size Estimate (startup / small / medium / large / XL)

• Resume Choice (SMB vs. Blend vs. AIncident vs. Enterprise). Each resume has the naming conventions written as Resume_<resume_choice>. Do not call the Resume Choice anything other than these naming conventions.

• Non-negotiables (list all that apply)

Analysis

1. Top 3 Alignments with the JD (must include evidence tags focusing on leadership impact and cross-functional influence).

2. Top 3 Gaps versus must-haves (prioritize the most critical gaps that could prevent progression, must include evidence tags or note unsupported).

Optimized Summary Rewrite

MANDATORY QUALITY CONTROLS - Perform these 3 manual checks:

1. JD-Specific Adaptation Check: Does this summary mirror 3+ specific keywords/phrases from the job description that differ from my current resume summary?

2. Unique Value Proposition Check: Does the first sentence create a distinct hook that would differentiate me from other candidates for THIS specific role?

3. Metrics Relevance Check: Are the 2-3 included metrics the most compelling ones for THIS job description’s priorities (not just my biggest numbers)?

Summary Requirements:

• Max 3 sentences. You are writing as a Senior Product Manager scoped for both mid-sized and enterprise tech companies.

• Cannot be 95% same as current Summary in Resume Choice selected. Your objective is to write a Summary that clearly hooks the reader that my experience solves their problem (in less than 5 seconds) that I need to be shortlisted (top 10% of candidates) for the role without over embellishment or lying about experience I don’t have.

• 1st sentence: “Senior/Seasoned <product/program> professional with X+ years leading…” (always state seniority and years).

• 2nd and 3rd sentences: Start with action verbs, mirror JD keywords exactly where possible, and front-load metrics and impact within the summary. Utilize 5 metrics max from my Professional Experience.

• First line must contain an impact hook (scale, complexity, or outcome).

• No “if not X, then Y” language, no em dashes, no objectives or forward-looking language.

• Maintain strategic, senior-level tone that demonstrates scalable thinking and hands-on execution capability.

Keyword Recommendations

• 5 singular word swaps strictly grounded in my actual experience but reframed to match JD language.

• Must include evidence tags referencing specific achievements or technical capabilities.

• Focus Adaptation by Resume Choice:

  • SMB Resume: Emphasize agility, rapid iteration, and direct customer impact

  • Blend Resume: Balance strategic execution with tactical delivery and cross-functional coordination

  • AIncident Resume: Expanded Blend resume with personal bootstrapped AI projects end-to-end. Still balances strategic execution with tactical delivery and cross-functional coordination

  • Enterprise Resume: Focus on strategic and execution impact, platform thinking, and organizational transformation

• Omit if unsupported.

Motivation & Gap Handling

• SKIP THIS ENTIRE SECTION if Application Recommendation = Don’t Bother

• Why the Domain? 3 sentences on why the industry or domain excites me, tying the company’s mission to my professional values and strategic outlook.

• No negative reference to Amazon.

• No metrics or specifics from my resume.

• Key Gap + Solution: 1 sentence naming the primary resume gap for this role and how my existing experience addresses it (must be evidence-backed, focusing on the most critical gap that could impact hiring decision).

Stylistic Rules and Parameters

Core Rules

• Use parentheses or commas for separation (no em dashes)

• Maximum two items per sentence in lists

• No direct resume metric copying without rewording

• All recommendations must be evidence-backed

• 15% AI polish, 85% humanization target

• Senior-level tone adaptation by Resume Choice:

  • SMB Resume: Strategic thinking with startup agility and hands-on execution focus

  • Blend Resume: Balanced strategic oversight and tactical execution across company stages

  • AIncident Resume: Balanced lean startup, strategic oversight, and tactical execution across company stages

  • Enterprise Resume: Platform thinking and enterprise-scale execution with systematic organizational impact

Banned AI Phrases

• “In today’s digital age/competitive landscape/rapidly evolving world”

• “In the realm/world of” and “When it comes to”

• “It’s worth noting/important to understand/recognize”

• “At the end of the day” and “In conclusion/To summarize” (mid-content)

• “Delve into” (use explore/examine), “Leverage” (use apply/use)

• “Navigate the landscape,” “Robust,” “Comprehensive,” “Vibrant,” “Dynamic”

Overused Transitions to Limit

• “However” (prefer but/though), “Furthermore,” “Moreover”

• “In addition,” “Additionally,” “On the other hand”

• “In contrast,” “Nevertheless”

Structural Patterns to Avoid

• Rule of three overuse (”adjective, adjective, and adjective”)

• Present participle endings for analysis (”-ing phrases”)

• Editorializing with vague significance claims

• Summary paragraphs restating main points

• Uniform formal tone throughout

Humanization Techniques

• Include contractions and personal qualifiers (”I think,” “probably,” “seems”)

• Vary sentence lengths dramatically (long complex + short punchy)

• Add minor imperfections or casual phrasing

• Use slightly awkward transitions between ideas

• Mix very short paragraphs with longer ones

Quality Requirements

• Cite specific studies, data, expert opinions

• Include concrete examples over abstractions

• Reference personal experience or observation

• Use industry terminology naturally

• Acknowledge complexity and uncertainty

When I say “New role:” the following text is the job description. Use only the inputs I provide.

Sample Analysis Request: Create 4 hypothetical job descriptions for analysis for SMB, Enterprise, AIncident, and Blend scenarios (1 each) to demonstrate the complete output format and ensure all parameters are working correctly.

---

What Drake Built Isn’t Just a Prompt, It’s a Mindset

Looking at Drake’s master prompt, you might think: “This is too complex. I could never build this.”

Drake thought the same thing at application 200.

But here’s what changed: He stopped treating his job search like a chore and started treating it like a product. He tracked metrics, ran experiments, built feedback loops, and iterated relentlessly.

The prompt was refined through 30+ iterations and 500 applications. It started as four messy chat windows and evolved into something that actually works.

Again - The Honest Truth About This Prompt

Don’t just copy and paste it.

This is Drake’s prompt - built for his background, his voice, and his job search goals.

Your version won’t look like this. Your experience is different. Your industry is different. Your voice is different.

Instead, use it as a blueprint. Study the structure: the integrity rules, the self-checks, the if/then logic, and the banned phrases. See how he grounded it in recruiter feedback and his own data.

Then build your own.

The power isn’t in Drake’s prompt; it’s in the process of building one that works for you.

Now It’s Your Turn

Here’s how to start:

1. Pick one pain point in your job search
Resume writing? LinkedIn outreach? Tracking applications? Start there.

2. Build a simple prompt to solve it
Don’t aim for perfection. Version 1 should be embarrassingly simple.

3. Track the results for 10-20 applications
Create your dataset. Every rejection is a training sample.

4. Refine based on what worked
Add guardrails. Test variations. Iterate like hell.

When Drake started, he didn’t have a master prompt, but he had frustration, Excel, and a willingness to iterate.

The Real Lesson

Job searching is a skill. AI doesn’t replace that skill - it amplifies it.

The job search isn’t getting easier. But you can get better at it.

Track. Test. Refine. Repeat.

That’s not AI magic—that’s product management applied to your career.

And if Drake can do it between rejection emails and interview prep, so can you.

Start with your data. Build your system. Ship your first version.

Then iterate like hell.

The next offer might be just 50 applications away.