How I AI“I’m incapable of doing my job without AI”: How this PM uses Claude + ChatGPT as his second brain
CHAPTERS
- 0:00 – 5:37
Project folders as “second brains” for PM context switching
Claire and Amir frame the core problem: PMs juggle constant context switching across initiatives. Amir introduces his workflow of maintaining separate Claude/ChatGPT “brains” (projects) filled with files, instructions, and ongoing threads to preserve context and speed up decisions.
- •Why context switching is uniquely painful for PMs
- •Using separate AI projects to silo initiatives and retain memory
- •What a “brain” looks like: files + instructions + long-running threads
- •How these setups make Amir feel “incapable” of working without AI
- 5:37 – 6:35
What to put in a PM brain: kickoff docs, PRDs, and “anything that’s text”
Amir explains how he seeds a project so the AI quickly understands the initiative. The key is to start with grounding references (kickoff decks, PRDs, artifacts) and then keep expanding the corpus as work progresses.
- •Start with reference artifacts that define goals and context
- •Seed the project with any available data, even rough inputs
- •Use a “ping-pong” back-and-forth to clarify what you’re trying to achieve
- •Continuously add new outputs back into the knowledge base as you learn
- 6:35 – 8:35
The agents problem: seeking unbiased, external customer truth
Amir shares why he turned to external conversations while leading an AI agents initiative at monday.com. With lots of internal opinions and fuzzy definitions, he wanted unbiased signals on what people actually expect from agents.
- •AI agents are high-hype with inconsistent definitions and expectations
- •Internal stakeholders often carry narratives and strong opinions
- •Goal: find a “voice of reason” outside the company
- •Reddit as a source of candid, unfiltered demand signals
- 8:35 – 12:01
Claude-assisted Reddit scraping: from idea to step-by-step implementation
Amir walks through how he used Claude to design a free/automated approach to gather online discussions, then narrowed to Reddit due to API accessibility. Claude provided an extremely granular, beginner-friendly setup guide that Amir followed end-to-end.
- •Start broad (“everything online”) then narrow to feasible sources
- •Why Twitter/LinkedIn are harder (API limits/paywalls)
- •Claude as a procedural tutor: terminal setup, packages, credentials
- •Building a script by filling in client ID/secret and search terms
- 12:01 – 12:36
From raw scrape to dataset: generating 34,000 rows of conversations
The scraper outputs a massive CSV representing Reddit threads relevant to monday.com, AI, and agents, including competitive comparisons. Amir emphasizes the value of scaling discovery beyond manual searching and collecting multiple themed files.
- •What the exported dataset looks like at scale (34,000 rows)
- •Expanding scope: competitor comparisons and adjacent themes
- •Turning scattered conversations into a reusable research asset
- •Using AI to avoid manual trawling while keeping breadth
- 12:36 – 13:32
Claude as analyst: summarizing themes with frequency, weights, and prioritization
Amir feeds the dataset back into Claude and asks for structured analysis that he can use to prioritize. He requests tables with frequency/percentages and key discussion points to translate qualitative chatter into decision-ready inputs.
- •Prompting for a table: themes, frequency, percentage, weights
- •Using quantified outputs to drive prioritization conversations
- •Extracting “top areas” people want from AI/agents
- •Transforming qualitative feedback into product strategy signals
- 13:32 – 14:23
Trust but verify: spot-checking and demanding quotes for fact-checking
Claire probes whether the analysis is reliable; Amir explains his validation workflow. He spot-checks with keyword searches and asks the model to include direct quotes so he can trace conclusions back to source rows.
- •Why you can’t manually verify 34,000 rows—but you can sample
- •Keyword searches as a quick sanity check
- •Asking for 1–2 reference quotes per claim/theme
- •Tracing quotes back to the CSV for validation
- 14:23 – 16:12
Building an effective knowledge base: PDFs from decks, websites, and support pages
They show how the Reddit files and internal artifacts are stored inside the project to become durable context. Claire highlights a practical trick: printing web pages (marketing, pricing, support) to PDF so the model has scoped, stable source material.
- •Adding scraped files back into the project’s knowledge store
- •Including kickoff decks and PRDs to anchor internal strategy
- •Printing websites/support/pricing pages as PDFs for clean ingestion
- •The mindset shift: “everything is text,” including slides
- 16:12 – 16:55
Instruction design: forcing pushback, candid feedback, and PM-grade reasoning
Amir describes how he writes system-like instructions so the AI behaves like a strong PM partner rather than a cheerleader. He explicitly requests challenge, professional standards, and strategic thinking so conversations stay sharp and useful.
- •Avoiding overly supportive AI responses by demanding pushback
- •Embedding expectations: product strategy and product sense mindset
- •Using the project as a continuous thought partner (“ping-pong”)
- •Accumulating thousands of threads as living context
- 16:55 – 18:28
Day-to-day PM execution: PRDs, narratives, and rapid stakeholder responses
Amir explains how these brains help him turn messy work into deliverables and respond quickly in the middle of the day. A common win is generating crisp, accurate summaries for marketing/comms when they need quick launch messaging.
- •Starting from scope → outline → narrative/PRD-ready output
- •Re-uploading drafts and docs to deepen the brain over time
- •Handling ad-hoc Slack asks (e.g., “two-line description”) fast
- •Reducing mental load while improving responsiveness
- 18:28 – 21:42
Custom writing coach GPT: training on Lenny/Wes Kao to shorten Slack messages
Amir shares a personalized GPT built from writing guidance (newsletters, books) to address recurring feedback: his writing is too long. The GPT rewrites messages to be concise while preserving a natural voice and avoiding telltale AI style.
- •Using performance feedback (“too long”) as a build target
- •Loading newsletters/books as the style and rules corpus
- •Explicit constraints: maintain voice, avoid over-bulleting/dashes
- •Practical workflow: draft → paste → rewrite → better responses
- 21:42 – 24:09
AI for professional development: turning feedback loops into daily practice
Claire and Amir discuss how AI makes skill-building actionable compared to relying on managers to edit work. They expand the idea to loading performance reviews and peer feedback into a GPT for reflection and blind-spot checks.
- •Why traditional coaching is slow: delayed manager feedback loops
- •AI as an always-available coach for repeated, small improvements
- •Using peer feedback/performance reviews to surface blind spots
- •Sharing custom GPTs with colleagues to scale benefits
- 24:09 – 31:49
Mock interviews with GPT voice mode: realistic product sense practice
Amir explains why voice mode changed his interview preparation: it’s interactive, challenging, and feels closer to a real interview than text. He demonstrates how to instruct voice mode to avoid hints, stay candid, and deliver feedback at the end.
- •Replacing passive video watching with active, spoken practice
- •Why text-based practice isn’t equivalent to verbal performance
- •Prompting voice mode: no leading, candid, end-of-interview feedback
- •Optionally loading CV + job description for tailored coaching
- 31:49 – 33:04
Additional voice-mode use cases and parenting/education ideas
Claire and Amir brainstorm adjacent uses for voice mode beyond interviews, including public speaking practice and reading coaching for kids. They note the broader potential (and complexity) of AI with children and learning routines.
- •Public speaking rehearsal and structured feedback via voice
- •Reading aloud practice with emphasis/punctuation coaching
- •AI + kids as a large emerging topic (future roundtable idea)
- •Voice interaction feels more natural than mirror practice for many
- 33:04 – 38:50
Recap, power-user advantage, and handling bad outputs (the ‘yell’ + exemplar method)
Claire recaps Amir’s full workflow: second brains, scraping + analysis, writing coach, and voice-mode interviews. Amir explains the key advantage—being “everywhere at once”—and closes with his tactics for correcting AI: emotional emphasis plus providing a concrete example of the desired output.
- •End-to-end workflow recap: collect → analyze → store → iterate
- •Power-user edge: rapid answers with deep context on demand
- •Bonus: Amir’s in-progress generalized scraper tool idea
- •Getting AI back on track: ‘this isn’t it’ + provide a target example
