How I AIClaude Code for product managers: research, writing, context libraries, custom to-do system, more
CHAPTERS
Teresa Torres’ background and why she’s all-in on Claude Code
Teresa Torres (Continuous Discovery Habits) explains how her AI usage evolved from web chat tools into Claude, then into Claude Code inside VS Code. The key shift is treating Claude like a pair-programming partner for non-coding work too.
- •Started with browser-based ChatGPT, moved to Claude for stronger writing help
- •Learned VS Code + Git after a project needed production-quality engineering practices
- •Claude Code in the terminal became a workflow “game changer”
- •Adopts a “pair programming” mindset for tasks, writing, and daily operations
Escaping Trello lock-in: building a personal task system with Claude
Teresa shares the motivation behind creating her own task management workflow: heavy note-taking and fear of losing or locking away her data in Trello. Moving tasks into a text-first system makes them searchable, portable, and AI-actionable.
- •Trello held both tasks and long-form thinking notes, creating lock-in and poor search
- •She evaluates each task: automate, augment, or keep it human
- •Putting tasks in Claude’s view enables: “What can you do for me today?”
- •AI becomes proactive—spotting opportunities to help, not just responding to prompts
Demo: the /today slash command that generates her daily plan
Teresa demonstrates a custom Claude Code slash command that compiles her day automatically. It creates a “today” file, pulls in due/overdue tasks, ongoing ideas, and research digest items—turning a daily ritual into a one-command workflow.
- •Uses a custom slash command (/today) with a detailed instruction prompt
- •Generates a daily “today file” she reviews every morning (even weekends)
- •Pulls from tasks due today, overdue tasks, and “ideas in progress”
- •Optionally checks Trello for team coordination items
Under the hood: Markdown + Obsidian as the task database
The system is built on plain Markdown files stored in Obsidian, where each task has structured YAML front matter. Claude (and scripts) can reliably query these files by due date, type, and tags, enabling fast automation without a GUI.
- •Tasks live as individual Markdown files in folders (Tasks/Ideas/Bugs, etc.)
- •Each task includes YAML front matter fields (type, due date, tags)
- •/today searches the Tasks folder for due dates and past-due items
- •A Python script supports the “search and assemble” behavior
Creating a new task in seconds (and why speed matters)
Teresa shows how she creates a new task by simply typing an instruction to Claude while working in the terminal. The benefit is avoiding slow, click-heavy interfaces and capturing tasks in the same environment where she’s already operating.
- •Creates tasks via natural language; Claude creates the file and sets due dates
- •Claude can update the “today list” after creating a task (with some caveats)
- •Avoids browser context switching and GUI steps like date pickers and labels
- •Keeps Claude open all day for rapid “bounce over” task capture
Claude-driven tagging and flexible views (e.g., sales pipeline on demand)
A major advantage of Teresa’s system is that Claude handles task tagging, making it easy to generate dynamic views like a sales pipeline without manual upkeep. She maintains a tag taxonomy with Claude to keep the system consistent over time.
- •Claude auto-suggests and applies tags, reducing manual task hygiene
- •Can ask: “What’s my sales pipeline?” and Claude compiles status from tagged tasks
- •Maintains a tag taxonomy in project-level instructions (Claude MD)
- •Iteratively refines tags when she sees inconsistencies
Embedding work notes inside tasks for powerful retrieval
Teresa keeps in-progress notes directly inside the task file, including bugs and partial work state. This creates a searchable “work journal,” and Claude helps recover details even when she only vaguely remembers what she wrote.
- •Notes live alongside tasks; bugs and context can be captured mid-task
- •Claude can locate information across tasks even with fuzzy memory or wrong wording
- •Traditional task tool search is often weak for embedded context
- •Claude acts as a local semantic search engine across her text corpus
Daily automated research digest: turning an academic firehose into a routine
Teresa explains a workflow that delivers a daily research digest into her task list. She scans titles/abstracts, manually downloads a few PDFs, and gets high-quality method-focused summaries the next day.
- •Tracks multiple research topics (e.g., synthetic users, collaboration, personas)
- •Daily arXiv/preprint search results appear in a Markdown digest
- •Manual selection of PDFs acts as an intentional filtering step
- •Next-day summaries help decide what’s truly worth deep reading
How the research plugin works: scripts, cron jobs, and summarization agents
She built the research system as a plugin backed by two Python scripts running on cron schedules. One gathers new papers from sources; the other detects newly saved PDFs and triggers Claude Code agents to produce structured summaries.
- •Plugin is in a public repo but still early-stage/testing
- •Morning script: searches arXiv daily; weekly script: searches Google Scholar Sundays
- •Tracks “already seen” vs. new papers and searches via a personal keyword config
- •Night script: detects new PDFs and triggers Claude Code to generate summaries
Summaries that emphasize rigor: methods, measures, and effect sizes
Teresa intentionally designed her summarization prompt to focus on the parts that matter for judging research quality. This lets her quickly spot weak measurement choices and decide whether a paper is credible or useful.
- •Summaries go beyond generic abstracts; emphasize methods and study quality
- •Includes details like effect sizes and measurement reliability (when available)
- •Helps her act as an editor for unreviewed preprints
- •Example: quickly critiqued a purchase-intent paper and wrote a high-performing LinkedIn post
Extending the idea: competitive analysis and other information digests
Beyond academic research, Teresa describes using Claude to “Google for her” and generate structured reports like competitive analyses. She also imagines similar digests for social content (e.g., LinkedIn), though API access is a constraint.
- •Blog post walkthrough: beginner-to-magic moment using Claude Code for competitive analysis
- •Claude can aggregate sources into price tables and feature comparisons
- •Desire: filtered LinkedIn-relevant posts without the normal UI (API limitations)
- •Core theme: digest + filter + operationalize overwhelming inbound information
Context libraries: building small, reusable files for ‘lazy prompting’
Teresa maintains an “LLM context” vault: many small, focused Markdown files covering business and personal context, plus a writing style guide. She learned that too much context hurts performance, so she uses indexes to load only what’s needed.
- •Creates context iteratively: “What did we learn today that should be documented?”
- •Claude co-authors context files (e.g., long-form writing style guide)
- •Uses index/profile files as a map to available context
- •Prefers many small files over one giant document to avoid irrelevant context pollution
Claude as a writing partner: research checks, critique, hooks, and typos
Teresa doesn’t want to automate her writing, but uses Claude heavily for augmentation: validating claims, strengthening hooks, critiquing sections, and cleaning up errors. The custom style guide makes feedback specific and aligned to her voice.
- •Uses Claude to fact-check and find evidence while she keeps writing
- •Requests targeted critique: hook strength, section clarity, what’s working/not
- •Style guide enables non-generic feedback tailored to her audience and goals
- •Occasional LLM-heavy drafts (e.g., transcript-to-stories, podcast theme synthesis), with transparency
Lightning round: tool stack, feature wishes, and what to do when Claude gets stuck
Teresa shares she mostly sticks with Claude Code + VS Code rather than chasing new tools like Cursor. When Claude derails, she resets with /clear and relies on documentation/context files to restart without re-explaining everything.
- •Prefers adopting new tools only when there’s a clear gap
- •Uses ChatGPT in-browser occasionally for convenience; uses Descript for editing
- •Wishes: better LinkedIn access and better text-to-image typography/quotes
- •When stuck: /clear, start fresh, and lean on documented context for continuity
