Teresa Torres, author of Continuous Discovery Habits, proves you don't need to be an engineer to build production-quality automation. Her Claude Code + Obsidian setup handles task management, daily research digests from academic papers, and writing workflows - all without traditional dev experience. This episode shows what's possible when Claude Code meets methodical file structure.

This is a very similar set up to what I run and covered in my Claude Code for PM’s series. You can dig in for a more detailed step by step walkthrough here: Part 1: Why Every PM Needs Claude Code

💡 Top tip - Read the TL;DR below to understand Teresa's complete setup, then dive into the task management or research automation sections that match your biggest pain point. The workflows are reusable regardless of your role.

Key insights from the full article:

Teresa's setup starts with something deceptively simple: organised folders in Obsidian. Every research topic gets a directory with two subdirectories - sources/ for PDFs and notes/ for summaries.

This mirrors the three-layer memory architecture covered in the Claude Code series Part 3: Global Layer (persistent preferences), Project Layer (topic-specific context), Reference Layer (on-demand documents). Teresa's implementation is pure Reference Layer - no complex .claude/ configuration files, just clean folder hierarchies.

Why this works:

When Claude Code searches your vault, structured folders create natural context boundaries. Teresa's topics/synthetic-users/sources/ folder tells Claude "these are academic papers about AI doing user research." When she asks Claude to summarise a paper, it understands the surrounding research landscape automatically.

The contrast with the Claude Code series: Part 3 described a vault with Plans/, Notes/, Product-Tapas/ - domain-driven top-level folders. Teresa's approach is topic-driven subdirectories. Both work because they create searchable hierarchies Claude can navigate.

Key Takeaways:

Teresa moved from Trello to Obsidian markdown for a surprising reason: data sovereignty. "I was worried - how am I ever going to get my data out of Trello?"

The solution: plain markdown to-do lists in Obsidian. Now Claude can read her task list directly and answer questions like:

The "pair task management" pattern:

This is lighter-weight than custom slash commands but still leverages Claude Code's ability to read your vault and execute actions.

Key Takeaways:

Teresa's killer workflow solves a problem familiar to anyone tracking industry research: too many papers, not enough time.

The system:

The breakthrough moment:

Teresa saw a paper on "purchase intent" in her daily digest, had Claude summarise it, immediately spotted a methodological flaw (unreliable survey instrument). The next day, Ethan Mollick shared the same paper on LinkedIn. Teresa wrote a detailed critical review and published it - becoming one of the first voices to question the study's validity.

Why this worked: The summarisation skill focuses on methods and effect sizes - the details academics use to evaluate research quality. Teresa didn't just get "here's what the paper says" summaries; she got "here's whether you should trust it" analysis.

Comparative insight: My Claude Code series Part 3 described 10 VCA research commands (/competitor-snapshot, /find-reddit-threads, /find-linkedin-posts etc.) - outward-facing research for competitive intelligence and market trends.

Teresa's research workflow is inward-facing - building a personal knowledge base from academic sources. The pattern is identical: automate discovery, let Claude synthesise, store results in searchable markdown.

Key Takeaways:

Teresa describes a workflow that feels like magic: fuzzy search with context awareness.

Example:

Teresa: "Hey Claude, I have a thing called 'new blog post tomorrow'"

Claude: "I can't find anything called 'new blog post tomorrow', but I have this thing that says 'article Wednesday.' Is that what you're looking for?"

Teresa: "Whoa, Claude, that is what I'm looking for."

What's happening here:

Claude Code reads the entire vault structure, understands semantic relationships (blog post ≈ article, tomorrow ≈ Wednesday if today is Tuesday), and tries every permutation of searches until it finds the match.

Obsidian's Auto-Link plugin for automatic connections between notes helps this even more helping build explicit connections through wiki-links, making fuzzy search even faster.

Both patterns work because:

Key Takeaways:

Teresa rarely writes LLM-first (generating drafts with AI). Instead, she pairs with Claude while writing - similar to how engineers pair-program.

The workflow:

What Claude does NOT do:

Comparative insight: My Claude Code series described newsletter automation - Gmail → Claude → Obsidian → Notion → Beehiiv draft. That's LLM-first content creation (AI generates summaries, human edits).

Teresa's approach is human-first content creation (human writes, AI assists). The series' automation saves 5-10 hours/week by removing mechanical work. Teresa's pair-writing preserves creative control while removing friction (typos, research lookups, fact-checking). Ho

Why pair-writing works for thought leadership:

AI-generated content has a recognisable feel - generic, over-explained, lacking edge. Teresa's LinkedIn posts and blog articles maintain her distinct voice because she owns the argument, Claude handles the mechanics.

The "fancy nails" story:

Claire Vo jokes that she now wears fancy nails despite being a terrible typist - because Claude fixes her typos automatically. This is friction removal in action. Small improvements (faster typo correction) enable new behaviors (caring less about typing accuracy).

Key Takeaways:

Teresa's path to Claude Code wasn't overnight. She describes a gradual evolution through four distinct phases:

Phase 1: ChatGPT in Browser (Like Everyone Else)

Started with basic prompts in the ChatGPT web interface. No integrations, no file access, purely conversational.

Phase 2: Claude for Writing

Switched to Claude because it's "a little bit better writer." Used Claude.ai web interface for content generation and editing.

Phase 3: VS Code + Real Engineering

Teresa had been coding in the AWS Management Console (no version control, no IDE) for four years. Her husband begged her to use an IDE.

The forcing function: a project she built was being integrated into a production product. She needed to level up her engineering practices. This meant learning Git, using VS Code, and becoming "a real engineer."

Phase 4: Claude Code Everywhere

Once Claude Code was integrated into VS Code, Teresa discovered the pair-programming model. "Engineers pair-program with Claude when they use Claude Code. I pair-program now with everything I do, even if it's not programming. I pair task-manage, I pair-write, I pair-everything."

I have a three-tool stack: Wispr Flow (voice input), Claude Code (orchestration), Obsidian (knowledge base).

Teresa's stack is slightly simpler: Claude Code + Obsidian. No voice input, just two tools used methodically.

The lesson: You don't need the complete stack to get transformational value. Teresa achieves similar productivity gains (task automation, research synthesis, writing assistance) with fewer moving parts.

Key Takeaways:

Teresa's most relatable moment: "I was writing my code in the AWS Management Console. I literally had no version control."

Her husband spent four years telling her to use an IDE. She resisted until she needed to contribute code to a production system - then the pain of not having Git, diffs, and proper tooling became unbearable.

Why this matters for PMs and non-engineers:

You don't need to be a "real engineer" to benefit from engineering tools. Teresa's research automation, task management, and writing workflows all run in terminal + Obsidian - tools that work for anyone who can write markdown.

The mental shift: Once you have Claude Code reading your files, you can describe what you want in plain English instead of writing code. Teresa's research digest automation didn't require her to learn arXiv's API - she told Claude "fetch papers about X daily and summarise them" and Claude figured out the implementation.

I covered setup instructions for Obsidian, Claude Code, and optional tools (n8n, MCP servers, API connections) in the Claude Code series. The guide assumed readers would install everything and build automation from day one.

Teresa's journey shows you can start with just Claude Code + Obsidian and add complexity only when you hit a specific gap. She added research automation after establishing her task management workflow. Incremental adoption, not big-bang transformation.

Key Takeaways:

Teresa is refreshingly honest about where AI falls short: "I hate AI-generated content. Reading other people's AI-generated comments kind of breaks my soul a little bit."

This is why she pair-writes instead of generating drafts. AI content has a recognisable feel - verbose, generic, lacking opinion. Teresa's LinkedIn presence depends on her distinct voice and spicy takes (like the critical paper review that went viral). AI can't replicate that.

The balance:

Comparative insight: The Claude Code series Part 3 described newsletter automation that generates "Not Boring" style content. That workflow is 60% AI-summarised, 40% human-edited. Teresa's workflow is 80% human-written, 20% AI-assisted.

Both are valid. The series' approach scales content production (5-10 hours saved weekly). Teresa's approach preserves thought leadership positioning (no one mistakes her writing for generic AI output).

The principle: Match AI involvement to the strategic value of the output. Newsletter summaries can be AI-heavy (readers want signal, not distinctive voice). LinkedIn thought leadership must be human-first (readers follow Teresa for her takes, not generic wisdom).

Key Takeaways:

Don't try to replicate Teresa's entire system. Start with one workflow that solves your biggest pain point:

If task management is chaos:

If research overwhelms you:

If writing feels slow:

If you're stuck in web interfaces:

The compounding effect: Each workflow builds on the previous. Task management in Obsidian → Research papers saved to Obsidian → Claude searches both when helping with writing. The system gets smarter as you feed it more structured content.

Key Takeaways:

Teresa and Claire agree: LinkedIn needs an MCP server (Model Context Protocol integration). Teresa wants to read LinkedIn content in her terminal without hitting the paywall or seeing AI-generated comments.

Why this matters:

LinkedIn is where thought leaders publish, but the web experience is hostile to productivity - ads, algorithmic feed, login walls, notifications. An API-backed terminal experience would let you:

Comparative insight: The Claude Code series Part 3 described 10 VCA research commands including /find-linkedin-posts which uses WebSearch to find high-engagement posts. That's a workaround for the lack of direct API access.

Teresa's request is first-party integration - LinkedIn building an official MCP server so Claude Code can read posts, comments, and profiles directly. This would eliminate the WebSearch workaround and provide real-time, structured data. But I do wonder how much more AI slop we would have to wade through to get to the stuff we want to read…`

The bigger pattern: As more tools build MCP servers, Claude Code becomes a universal interface for your entire toolstack. Teresa's already using it for task management, research, and writing. LinkedIn integration would add professional networking to that list.

Key Takeaways:

Teresa Torres proves you don't need to be a "real engineer" to build production-quality automation. Her setup - Claude Code + Obsidian + methodical file structure - delivers:

✅ Task management that sees your priorities and delegates mechanical work ✅ Research automation that builds searchable knowledge bases from academic papers ✅ Pair-writing workflows that preserve your voice while removing friction ✅ Local search that finds content even when you misremember filenames

Just two things:

I showed you can go much further - Part 3 of my series showed a maximalist setup (40+ automation components, newsletter pipelines, custom commands). Teresa shows a minimalist alternative that achieves great productivity gains with fewer moving parts.

Both approaches work because:

Pick your starting point based on pain point severity.

The compounding effect works either way: each workflow feeds the next, Claude gets smarter as your vault grows, productivity gains multiply over time.

Key Takeaways:

Related Resources:

📖 Product Tapas Claude Code Series:

📚 Teresa Torres:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.