CHAPTERS
- 0:00 – 1:41
AI for leaders: decomposition as the core skill (and why Claude simplifies poorly framed problems)
Matt sets the central thesis: AI is powerful at solving defined problems, but leaders create leverage by decomposing messy problems into clear parts. Without that decomposition, models flatten nuance and produce plausible-but-misaligned output.
- •AI performance depends on the quality of problem decomposition
- •LLMs tend to simplify the problem space unless guided deliberately
- •Leadership value comes from dwelling on the problem long enough to separate the pieces
- •This mindset frames the rest of the episode’s workflows
- 1:41 – 3:50
Why most “AI for PMs” advice misses leadership work (and what this episode will teach)
Aakash explains that most AI content targets individual-contributor PM tasks (PRDs, feature analysis) rather than leadership deliverables. He previews the three outcomes: building an all-hands fast, running metrics retros, and Matt’s weekly AI stack.
- •Gap: IC-level prompts vs company/board-facing artifacts
- •Leadership outputs: all-hands, C-suite retros, strategy narratives
- •Promise: 80/20 split—what to delegate to Claude vs what to own
- •Focus on real documents, not hypothetical examples
- 3:50 – 5:10
The 5AM all-hands build: goals, audience, and narrative shape before any prompting
Matt recounts building his portion of a Q2 roadmap all-hands in a single morning using a slide template plus Claude. He starts with fundamentals—audience and takeaway—then decides the narrative arc himself (acknowledge the recent launch, bridge to what’s next).
- •Know your audience mix (GTM + engineering) and desired takeaway
- •Start from story craft: “where did we leave our hero?”
- •Don’t outsource narrative structure; leaders set the shape of the story
- •Use templates for visuals, reserve judgment and sequencing for yourself
- 5:10 – 6:25
Take inventory before you open Claude: collecting the best raw materials
Instead of jumping into slide creation, Matt first inventories inputs that Claude can transform. Prior internal materials—especially demo day recordings—become raw feedstock, even if they’re misaligned to the all-hands audience.
- •Treat Claude as a transformer of raw materials, not a mind-reader
- •Identify the highest-signal sources already available in the company
- •Demo day is rich input but needs reshaping for non-engineering audiences
- •Start with inputs, then decide transformations
- 6:25 – 9:44
Using Zoom transcripts to extract screenshots fast (timestamps as the real deliverable)
Matt shows a pragmatic use of Zoom recordings: paste the transcript and have Claude extract timestamps for key moments so he can grab screenshots quickly. The goal isn’t “summarize the meeting,” it’s to reduce the time cost of hunting in video.
- •You can paste large transcripts, but give a narrow task with clear constraints
- •Ask for timestamps aligned to specific screenshot needs
- •Use Claude to avoid scrubbing a 50-minute recording manually
- •This enables “show, not tell” slides with real visual evidence
- 9:44 – 12:14
Sponsor break: evaluating replay-tool AIs with a rubric (LogRocket Galileo vs PostHog Max)
Aakash shares an eval-style comparison of two session replay AIs using the same questions and traces, scoring them on correctness, completeness, and hallucinations. He argues that confidence isn’t correctness, and instrumentation defaults (like exception capture) can dominate outcomes.
- •Set up: same site, same sessions, same questions; score with a rubric
- •Galileo scored higher largely due to auto exception capture
- •Example insight: identifying React errors breaking key buttons for many users
- •Takeaway: treat AI tooling as an eval problem, not a vibes problem
- 12:14 – 13:50
“Matrix multiplication”: pivoting raw demo-day content into the company’s strategic themes
Matt combines two sources—Zoom demo-day content and the strategy doc’s investment themes—and asks Claude to reorganize the raw material into the strategic shape leadership needs. He describes this as a pivot/transform operation that makes the content immediately usable for slides.
- •Provide both raw input (demo day) and target schema (strategy themes)
- •Ask Claude to categorize initiatives by annual investment themes
- •Think in transformations: source → target format (pivoting content)
- •The ‘click’ moment: now the material is shaped for executive storytelling
- 13:50 – 14:57
Build slides first, then generate the talk track from slide screenshots
Matt prefers to create the visual story first (what people should see), then have Claude draft a talk track that adds context rather than repeating bullets. He feeds screenshots of finished slides back into the same session to leverage accumulated context.
- •Order matters: visuals first to ensure “show, not tell”
- •Feed slide screenshots back to Claude to draft speaker notes
- •Instruct Claude to avoid regurgitating slide text; use broader context
- •Talk track polishing can take as long as slide creation
- 14:57 – 19:02
The ‘eager junior’ problem: stopping Claude from racing ahead into slop
Matt explains that Claude behaves like a talented but overly eager junior employee—rushing to deliver a polished artifact without enough clarification. His fix is to manage the session tightly: stop premature recommendations, drip context iteratively, and delay the final deliverable until the model has earned it.
- •Common failure: Claude jumps to a full deliverable (e.g., a Word doc) too early
- •‘Micro-hallucinations’: wrong form factor or workflow assumptions, not fake facts
- •Use firm guardrails: “Stop recommending next steps; wait for my instruction”
- •Prefer long iterative sessions over quick drafts plus endless revisions
- 19:02 – 22:56
Metaphors as a control system: starting abstract to force clarifying questions
On a walk, Matt uses a biology/water-cycle metaphor to design a customer lifecycle model without letting the model latch onto Customer.io specifics. Keeping it abstract makes Claude ask clarifying questions—behaving more like a senior partner—and yields a cleaner, transferable mental model.
- •Begin with analogy to avoid triggering premature domain assumptions
- •Abstraction encourages clarifying questions instead of eager execution
- •Iteratively define a lifecycle model (e.g., stages in a 2x2)
- •A clean model can later be stress-tested across domains, including your own
- 22:56 – 25:44
The game-night rule: layer complexity gradually so both you and the model stay oriented
Matt compares prompting to explaining a board game: start with core rules, then add exceptions and nuance only after the basics are stable. This prevents cognitive overload, keeps the thread coherent, and gives the leader time to refine their own thinking.
- •Start simple: objectives and core rules first
- •Add exceptions only after the foundation is locked in
- •Avoid dumping complexity all at once (model ‘indigestion’ and lost threads)
- •The process also improves the leader’s clarity, not just the AI’s output
- 25:44 – 38:17
Reveal the domain only when the model is clean: avoiding jargon and misreading the room
Matt waits to disclose the real business context until the abstraction yields a robust model, then applies it to Customer.io. He highlights a key limitation: political/room-reading calibration—Claude will reuse cute internal labels or loaded terms unless explicitly coached for the audience.
- •Delay domain reveal to prevent the model from ‘shortcutting’ to assumptions
- •When revealed, the model gets excited—so you must reassert constraints
- •AI often mishandles jargon and audience sensitivity (social/political calibration)
- •Persuasion depends on reader context (fatigue, recent events), which AI can’t reliably infer
- 38:17 – 40:34
Why AI-generated ‘executive alignment’ decks backfire (and what alignment actually requires)
Matt argues that using AI to generate alignment is likely to fail because executives are expert noise filters and can detect slop quickly. The outcome varies by hierarchy level—forced nods at the top, outright ignoring at lower levels—but either way true alignment doesn’t happen.
- •Alignment definitions vary by culture (disagree-and-commit vs shared consciousness)
- •Executives filter noise and detect BS/slop quickly
- •Hierarchy changes symptoms: polite nods vs being ignored entirely
- •Better approach: bring novel thinking and decomposed nuance, not flat AI outputs
- 40:34 – 46:52
Matt’s weekly AI stack: Claude for craft + Slack agents for analysis, scanning, and doc audits
Matt outlines a practical operating system: Claude desktop for deep transformations and writing, and Slack-based agents for day-to-day leadership leverage. Key tools include a Snowflake-connected analysis bot, an org-wide conversation scanner for product involvement, and ‘Chiefy’ to audit documents against a gold-standard corpus.
- •Slack analysis bot: natural-language queries over Snowflake with human verification
- •Scanner bot: monitors channels to flag where PMs should engage (remote/async scale)
- •Chiefy: audits new docs vs ratified company docs; finds discrepancies and staleness
- •Emphasis: AI keeps leaders close to ground truth while reducing monitoring burden
- 46:52 – 50:16
How Customer.io enables secure AI experimentation (and how to replicate the setup)
Matt explains that the stack works because leadership intentionally funds and supports controlled experimentation: internal agents, secure hosting, Slack access, and feedback loops. He suggests other companies replicate the approach with budget, guardrails, and leaders who model usage.
- •Enablement model: allow experimentation in a controlled, secure way
- •Support internal builders with hosting, security guidance, and budget
- •Integrate agents into Slack workflows for adoption and leverage
- •Lead by example: leaders use, maintain, and improve the tools
