Aakash GuptaComplete Course: Claude for PMs (Cowork + Code + Dispatch)
CHAPTERS
- 0:00 – 2:51
Why web chat is obsolete: Claude’s real power is in agentic surfaces
Aakash and Pawel open with a provocative claim: there’s little reason to use Claude as a plain web chatbot anymore. Pawel explains how real PM work quickly hits chat limitations—continuity, mobility, tool access, and switching contexts—making Cowork, Code, and Dispatch the better default.
- •Chat is fine for quick one-offs (grammar checks), but breaks for ongoing work
- •Hard limitations: can’t truly continue sessions across devices; context re-entry costs time
- •As tasks evolve (files → code → visuals → email), chat forces tool/context resets
- •Agentic interfaces (Cowork/Code/Dispatch) let you move seamlessly between workflows
- 2:51 – 7:46
What Anthropic’s shipping velocity signals about the future PM role
Pawel uses Anthropic’s rapid release cadence to illustrate a broader shift: companies are redesigning workflows around AI capabilities rather than simply inserting AI into old processes. The PM role moves toward ‘super IC’ territory—closer to engineering, design, and revenue-driving strategy.
- •AI enables workflow redesign, not just automation of existing steps
- •PM/PMM/designer/engineer responsibilities are converging via automation
- •PMs must grow technical comfort (e.g., terminal/engineering tools)
- •Strategy and revenue linkage become more central to PM effectiveness
- 7:46 – 10:17
Sponsor segment: tracing, evals, and the ‘trace → evaluate → fix’ loop with Arize
Aakash shares a concrete pain point from building agents: systems can appear fine until users surface hallucinations and wrong tool choices. He demonstrates how instrumentation and evaluation (with Arize + Claude Code) makes agent behavior observable and improvable.
- •Agent failures often come from lack of evaluation, not just prompting
- •Tracing reveals step-by-step tool calls and decisions
- •Claude Code can help propose evaluation criteria and implement fixes
- •Fast iteration loop can materially reduce error rates (e.g., 12% → <2%)
- 10:17 – 15:01
Tool map for PMs: when to use Chat vs Cowork vs Claude Code
Pawel provides a practical ‘if-this-then-that’ mental model for Claude’s surfaces. Cowork is for real files and multi-step workflows; Code is for codebases and local execution; chat is limited to basic Q&A and lightweight outputs.
- •Chat: basic drafting/summarizing; limited continuity and complexity
- •Cowork: works with real files, executes multi-step plans, can parallelize via sub-agents
- •Claude Code: optimized for codebases, debugging, databases, and local system commands
- •PMs can’t skip Code long-term because engineering work and artifacts are code-centric
- 15:01 – 21:04
Cowork power-user setup: projects, permissions, and a live file-automation demo
Pawel shows how Cowork is organized around folders/projects and demonstrates an end-to-end workflow: organizing invoices by month, removing duplicates, and handling mixed media. This illustrates why Cowork outclasses chat for day-to-day operational tasks.
- •Cowork projects/folders create reusable workspaces with tailored instructions
- •Granting folder access enables real file manipulation (PDFs/images)
- •Agent plans steps explicitly (extract dates, dedupe via hashing, create folders, move files)
- •The same folder can become a repeatable ‘inbox’ for ongoing automation
- 21:04 – 24:48
Skills + MCP connectors: progressive disclosure, Gmail/Slack workflows, and safer drafting
The discussion moves to ‘skills’ (procedures Claude can load on demand) and connectors (MCP servers) that integrate external tools like Gmail and Slack. Pawel highlights progressive disclosure to avoid bloated context, and he emphasizes drafting responses with human review rather than auto-sending.
- •Skills are modular procedures; Claude loads full instructions only when relevant
- •Connectors/MCP are the ‘USB for agents’—Drive, Gmail, Slack, and more
- •Inbox processing: Claude drafts replies; Pawel reviews and sends manually
- •Claude learns from approved edits over time, improving future drafts
- 24:48 – 30:59
PM Skills Marketplace: plugins, workflows, and forcing the right skill via slash commands
Pawel tours his viral PM skills repository and explains how plugins bundle domain skills (discovery, strategy, GTM, analytics). He shows installing the marketplace and demonstrates why explicit slash commands can override Claude’s generic training-data instincts.
- •Plugins group skills by domain; workflows chain multiple skills end-to-end
- •Marketplace installation enables quick access to reusable PM playbooks
- •Slash commands help ensure the correct skill activates vs generic responses
- •Example workflow: needs → opportunities → ideation → assumptions → experiments
- 30:59 – 36:28
How to write great skill files: markdown structure, triggers, and embedded knowledge links
Pawel breaks down the anatomy of a skill file: name/description/when-to-use plus stepwise instructions and role framing. Skills can also point to deeper articles as supporting context, effectively blending execution prompts with curated reference material.
- •Skills are structured prompts in Markdown; can be created by describing a process to Claude
- •Front-matter sections help agents decide when to load a skill
- •Skills can define personas (e.g., devil’s advocate) and step-by-step procedures
- •Optional embedded links provide deeper domain context at runtime
- 36:28 – 40:51
Strategy Canvas demo: generating McKinsey-style PowerPoint with skills
Using the installed skills plus Anthropic’s PPTX capability, Pawel generates a Product Strategy Canvas deck for ‘Amazon 2.0’. Aakash emphasizes the new baseline: PMs can produce polished, structured slides quickly, with frameworks (North Star, guardrails) baked in via skills.
- •Cowork can generate full PPTX decks with strong layout, icons, and structure
- •Skills enforce strategic frameworks (North Star, guardrails, trade-offs, metrics)
- •Output quality is approaching consulting-style deliverables in minutes
- •Takeaway: presentation quality should no longer be a bottleneck for PMs
- 40:51 – 44:33
Why PMs need Claude Code: navigating complex repos, reusable components, and web data
Pawel explains Cowork’s limitations for large hierarchies and codebases (no true explorer view) and why Code’s project structure matters. He shows examples of Claude-generated graphics and discusses how his system reuses components and automates research-heavy tasks like sourcing images/data from X.
- •Cowork is awkward for large file trees and codebases; Code provides the explorer/IDE model
- •Claude Code supports building reusable HTML/component libraries for infographics
- •Research automation: gather data from X via free FX endpoints or paid API wrappers
- •Non-coders can still use Code by conversing and asking the system to implement scripts
- 44:33 – 56:16
Building a ‘second brain for agents’: self-improving knowledge from posts and infographics
Pawel details a knowledge system inspired by Karpathy: instead of a personal wiki for humans, he curates a growing knowledge base for agents. The system converts observed content patterns into rules and hypotheses, tracks what’s confirmed vs rejected, and applies the learnings to future outputs.
- •Curate inputs (screenshots, posts, infographics) and have agents extract patterns
- •Store patterns as confirmed rules vs hypotheses awaiting more evidence
- •Maintain ‘rejected beliefs’ to prevent repeating disproven heuristics
- •Use the database to shape style, hooks, and formats while keeping human takes central
- 56:16 – 1:01:40
Claude.md and folder architecture: keeping context lean with indexed domain files
They cover a common mistake: stuffing everything into Claude.md until it bloats the context window. Pawel shows a lean Claude.md that acts as a router—explaining the project, where knowledge lives, and the workflows agents must follow—while detailed rules sit in dedicated domain files.
- •Claude.md should define purpose + navigation, not all detailed instructions
- •Store knowledge by domain in separate files with an index for routing
- •Workflows specify when to analyze visuals (token savings) and how to update hypotheses
- •Claude can write and maintain these instructions based on conversational guidance
- 1:01:40 – 1:05:22
Minimum viable self-improving setup: rules/hypotheses loop any PM can use
Aakash pushes for a PM-friendly starting point: a simple instruction set that makes Claude review existing domain rules before acting, apply confirmed rules by default, and update hypotheses as new examples arrive. The same pattern works for testing, release notes, hiring, pricing, or discovery—not just content creation.
- •MV prompt: review domain rules/hypotheses → apply rules → update knowledge after outcomes
- •Feed examples of good vs bad artifacts to let Claude infer patterns
- •Organize by domains (strategy, testing, marketing, pricing) to keep learning targeted
- •This replaces ‘prompt-from-scratch’ with compounding organizational memory
- 1:05:22 – 1:09:47
Browser automation: why Chrome MCP is token-expensive and when to use Agent Browser
Pawel explains why he stopped using Chrome MCP/Claude-in-Chrome for repeated workflows: screenshot-based browsing burns tokens fast. He recommends Vercel’s Agent Browser for headless, token-efficient interaction when no API/MCP integration exists (e.g., legacy tools).
- •Screenshot-driven browser control can be extremely expensive with larger models
- •Agent Browser renders headlessly and returns structured page affordances (buttons/text)
- •Use when you need data/actions from sites without APIs or MCP connectors
- •Applicable to legacy enterprise web apps (old CRM/SAP-like systems)
- 1:09:47 – 1:20:12
Dispatch + Web Sessions + remote surfaces: running agents from phone, anywhere
The conversation shifts to remote work: Dispatch as a walkie-talkie for background tasks, Web Sessions for cloud-hosted coding sessions, and why Channels didn’t add value for Pawel. Key benefits include multi-threaded work while mobile, and resilience when a laptop is offline by syncing projects to GitHub and running on Anthropic’s servers.
- •Dispatch: start multiple background tasks and review results asynchronously (mobile-friendly)
- •Web Sessions: cloud-hosted Claude Code sessions tied to GitHub repos; works even when laptop is offline
- •Managing complexity: use Code sessions when many parallel threads become hard to juggle in Dispatch
- •Remote workflow enables ‘life-integrated’ work—short feedback loops throughout the day
- 1:20:12 – 1:32:41
Common PM mistakes, what’s next for AI PMs, and n8n vs Claude Code for production
Pawel’s biggest critique: PMs repeatedly prompt from scratch instead of building a learning system. They forecast ‘super IC’ PMs orchestrating agents, and end with a pragmatic point: Claude Code is great for personal automation, but production workflows still need hard-coded guardrails—often via tools like n8n or API-based implementations.
- •Biggest mistake: chasing better prompts instead of organizing knowledge and learning loops
- •Near-term PM future: orchestrating multiple agents; less trivial work, more judgment and synthesis
- •Everything in Claude’s agent ecosystem is still underappreciated (under-hyped)
- •n8n remains relevant for production-grade automation with explicit rules, retries, and safety controls
