How I AIClaude Code for product managers: research, writing, context libraries, custom to-do system, more
CHAPTERS
- 0:00 – 2:43
Claude Code as a “pairing” companion for daily work (Teresa’s why)
Teresa explains how Claude Code evolved into her go-to productivity partner, not just for coding but for “pair programming” everything—task management, writing, and research. The core promise: move work into a context Claude can see, so it can proactively help and automate pieces of your day.
- •Motivation: getting notes out of Trello and making them searchable/portable
- •Shift in mindset: constantly asking whether AI should automate vs. augment a task
- •Claude’s advantage when it can ‘see’ your working files and task system
- •Pair-programming mental model applied beyond engineering
- 2:43 – 4:14
From web chat to terminal workflows: leveling up with VS Code + Git
Teresa describes her gradual transition from browser-based ChatGPT/Claude usage to terminal-based Claude Code inside VS Code. A real production integration pushed her to adopt more ‘real engineering’ practices (IDE, Git), which unlocked faster, tighter AI-assisted workflows.
- •Started in web chat; moved to Claude for stronger writing
- •Catalyst: a project moving toward production-quality integration
- •Adopting VS Code and Git as a turning point
- •Claude in the terminal becomes a daily, always-open work surface
- 4:14 – 6:07
Why build a custom to-do system: idiosyncratic workflows + escaping tool lock-in
Teresa argues task management is deeply personal, which is why off-the-shelf apps never fully fit—hence the explosion of to-do startups. Her own pain point was notes trapped in Trello, hard to search, and risky to export—so she rebuilt tasks as plain text that Claude can operate on.
- •Task management needs vary wildly; personal systems often win
- •Notes in Trello felt locked-in and not easily searchable
- •Text-first approach makes data durable and AI-accessible
- •Design goal: reduce UI friction and let Claude do the busywork
- 6:07 – 7:45
Demo: /today slash command that generates a daily plan from markdown tasks
Teresa shows her central routine: typing `/today` to generate a “Today” file. Claude scans tasks, identifies due/overdue items, pulls in ongoing ideas, and appends a daily research digest—creating a single operational dashboard for the day.
- •Slash commands as reusable shortcuts with detailed prompts
- •Daily output includes: Trello sync check, due today, overdue, in-progress ideas
- •Everything renders into a ‘today’ markdown file viewed in Obsidian
- •Research digest is integrated into the same daily workflow
- 7:45 – 8:38
How tasks are structured: Obsidian + YAML front matter + tags Claude maintains
The system relies on simple markdown files with consistent YAML front matter (type, due date, tags). Claude (and supporting scripts) searches the Tasks folder to assemble views, while also tagging work automatically so Teresa can query her workload by theme (e.g., sales pipeline).
- •Tasks are markdown files with YAML front matter (type/due/tags)
- •Claude searches by due date and status rather than relying on app UI views
- •Automatic tagging replaces unreliable manual tagging habits
- •Project-level ‘taxonomy’ of tags is maintained in Claude MD
- 8:38 – 10:55
Under the hood + creating tasks fast: Python scripts, minimal GUI, instant capture
Teresa demonstrates creating a new task conversationally, while Claude writes the task file and attempts to update the today view. A small Python script powers the “search due/overdue” logic; the real benefit is speed—capturing tasks without context switching to a browser UI.
- •Task creation: natural language → new markdown file in the Tasks folder
- •Python script supports searching tasks and generating today output
- •Avoids GUI friction: date pickers, moving cards, clicking through views
- •Claude is kept open all day; quick ‘bounce over’ task capture
- 10:55 – 14:12
Why Obsidian is in the loop: a unified file browser + AI-friendly text vault
Teresa wasn’t an Obsidian user initially, but adopted it for the tactile checklist feel and—more importantly—the file browser over a broader vault of work (tasks, writing, research, context files). The key is that everything stays in accessible markdown that Claude can read and manipulate.
- •Obsidian provides a practical file browser across many work domains
- •Markdown keeps all content legible to both humans and Claude
- •Unified vault includes tasks, research, writing, and context libraries
- •Enables dynamic queries like ‘what’s my sales pipeline?’ from tags
- 14:12 – 16:06
Task notes as searchable memory: logging work-in-progress details inside tasks
Teresa explains how embedding notes inside task files makes it easy to recover details later—especially when you recorded something ‘lazily’ and forgot where. Claude becomes a powerful local search engine, iterating through fuzzy matches until it finds the right context.
- •Capture bug notes and partial progress directly in the task file
- •Later retrieval: ask Claude even with imperfect recollection
- •Claude performs robust, iterative search beyond typical app search
- •Treats your task system as a searchable knowledge base
- 16:06 – 19:33
Daily research digest workflow: from arXiv searches to next-day paper summaries
Teresa walks through a research pipeline that turns an overwhelming stream of academic papers into a daily, actionable queue. Each day she reviews an arXiv digest, manually downloads a few PDFs, and then receives structured summaries the next day to decide what’s worth deep reading.
- •Daily arXiv search results saved into a markdown digest file
- •Manual selection step: download only the papers that look relevant
- •Folder structure per topic: sources (PDFs) + notes
- •Next-day summaries emphasize methods/effect size for quick evaluation
- 19:33 – 21:16
How the research plugin works: cron jobs + config keywords + Claude agents
Teresa explains the implementation: a public (but experimental) repo built with Claude’s help, powered by two Python scripts and cron jobs. One script runs searches (daily arXiv, weekly Google Scholar) and deduplicates results; the other detects new PDFs and triggers Claude Code agents to generate summaries.
- •Two-script architecture: search/dedupe + nightly summarization trigger
- •Cron jobs schedule the workflows on her computer
- •Config file of personally defined keywords drives queries
- •AI is ‘required’ mainly for summaries, but helped build the whole system
- 21:16 – 25:35
Filtering the firehose: why manual PDF selection still matters
Despite automation, Teresa keeps a human-in-the-loop filter: she downloads PDFs manually to avoid being overwhelmed. Spending 5–10 minutes scanning the digest preserves attention while still benefiting from next-day AI summaries.
- •Automatic downloading is possible but intentionally avoided
- •Manual triage prevents runaway volume and low-signal summaries
- •Time-boxed daily review keeps the system sustainable
- •Summaries arrive the next day without needing to ‘remember’ the task
- 25:35 – 32:57
Context libraries for ‘lazy prompting’: small files, indexes, and scoped instructions
Teresa shows her dedicated “LLM context” vault: many small, focused markdown files plus index/profile files that help Claude pull the right context on demand. She learned that stuffing everything into one Claude.md hurts performance—irrelevant context degrades output—so she uses scoped profiles (business vs. personal) and maps to relevant documents.
- •Iterative creation: ‘What did we learn today that we should document?’
- •Writing style guide co-created by Claude from her existing blog content
- •Index/profile files act as a map to available context documents
- •Principle: many small focused files > one giant document; scope matters
- 32:57 – 36:55
Claude as a writing partner: research checks, stronger hooks, style-aware critique
Teresa explains she rarely lets AI write first because she loves writing, but uses Claude heavily as an editor and collaborator. Claude fact-checks claims, suggests improvements (especially hooks and structure), applies her style guide, and cleans up typos—making the revision loop faster and more consistent.
- •Augmentation over automation for writing (with a few transparent exceptions)
- •On-demand research: ‘Is this claim true?’ while she keeps drafting
- •Targeted feedback: hook strength, section review, what’s working/not working
- •Style guide makes critiques specific instead of generic ‘AI feedback’
- 36:55 – 43:12
Lightning round: tool stack, feature wishes, and resetting with /clear
Teresa shares her minimalist approach to tools—she only adopts new products when there’s a clear gap. She mentions occasional ChatGPT browser use and Descript for editing, wishes for better LinkedIn access and reliable text-in-image quoting, and describes her go-to recovery move when Claude derails: `/clear` and start fresh, relying on documented context files instead of chat history.
- •Tool strategy: avoid overwhelm; try new tools only when current setup fails
- •Other tools: VS Code for coding; occasional ChatGPT; loves Descript
- •Wishes: LinkedIn access via API/MCP; better text-to-image typography/quotes
- •Failure mode handling: use `/clear` often; depend on reusable documentation/context
