How I AI: Teresa Torres's Claude Code System for Task Management, Automated Research, and 'Lazy' Prompting

In this episode of How I AI, Continuous Discovery Habits author Teresa Torres shows the local-first system she runs through Claude Code and Obsidian for daily tasks, academic research, and reusable context.

Teresa calls the approach "pair programming for everything." She works from the terminal whether she is managing tasks, writing, or researching, and keeps the underlying files on her own computer.

Her three workflows replace Trello with local task files, add a daily research digest to her to-do list, and organize enough reusable context that she can be "lazy with her prompting" without sacrificing quality.

Build a personal task manager with Claude Code

How to Build a Personalized Task Manager with Claude Code and Markdown: Build a task system from plain Markdown files, a small metadata schema, and a `/today` command that assembles overdue work, tasks due today, active ideas, and the research digest in one view.

How to Automate Academic Research with Claude Code and Python Scripts: Build a research digest that searches arXiv daily and Google Scholar weekly, filters results against your own topics, and turns selected PDFs into rigorous summaries of methods, findings, effect sizes, and limitations.

Teresa wanted her tasks in portable files rather than inside Trello. She built a task system in VS Code and Obsidian, with Claude Code creating and updating the files.

Run the /today command

Each morning, Teresa opens her terminal and runs:

/today

The custom command runs a Python script that scans her task files and creates a dated Markdown to-do list containing:

• Tasks due today: A checklist of items with today's due date.
• Overdue tasks: Every unfinished item whose due date has passed.
• In-progress ideas: Longer-term projects available after the day's required work.
• Research digest: A link to that day's automated research results.

Store each task as a Markdown file

Each task is a local Markdown file in an Obsidian vault. YAML front matter supplies structured metadata that Claude and other scripts can read.

A task file can look like this:

---
type: task
due_date: 2024-08-21
tags:
  - sales
  - course-launch
---
# Update sales page for the new course
- [ ] Draft new copy for the features section
- [ ] Find new testimonials
- [ ] Update pricing table

Notes live in the same file as the task. Teresa can record a bug or half-formed idea while working, then ask Claude to find it later without remembering the filename.

Manage tasks in natural language

Teresa adds a task by describing it in the Claude Code window she already has open.

During the conversation, she created one with this instruction:

new task, send. Thank you to Claire. do today. How I AI was a blast.

Claude creates the Markdown file, adds the due date and tags in YAML front matter, and includes the task in the daily to-do list. A project-level claude.md file defines the tag taxonomy.

Automate an academic research digest

Teresa wanted a regular view of new research without repeating the same searches. She connected a research plugin to the task system.

Review the daily digest

The /today output links to a Markdown digest of papers related to Teresa's configured topics, including synthetic users, team collaboration, and education.

Teresa scans the list and manually saves a promising PDF to the relevant research folder. That selection step is a useful filter against information overload.

Run the search and summary scripts

Two Python scripts run as cron jobs on her computer:

1. Search script: Each morning it queries arXiv, and on Sundays it also queries Google Scholar. A configuration file supplies Teresa's topics and records papers already shown.
2. Summary script: Each night it scans the research folders for newly saved PDFs and asks a Claude Code agent to summarize them.

Summarize the evidence Teresa evaluates

Teresa's summary skill asks for methodology, effect size, and other evidence she uses to assess an academic paper rather than returning only its abstract.

The summaries appear in the next digest. Teresa used one to examine a new paper, identify a methodological flaw, and write a critical LinkedIn review that became one of her best-performing posts.

Prompt less with an indexed context library

How to Create a Granular Context Library for 'Lazy Prompting' with AI: Give Claude a useful map of your work without loading everything at once. Store focused context in small Markdown files, connect them with profile indexes, and let short prompts pull only the files relevant to the task.

Teresa wanted short prompts without loading every detail of her work and life into every request. She split the context into small files and indexes instead of maintaining one large claude.md file.

Avoid one context file for every task

A single file made unrelated information travel together. A personal question about her dog did not need her business profile, marketing strategy, and product details.

Build a small, indexed file library

Teresa keeps focused Markdown files in an Obsidian vault called LLM Context. Business and personal folders separate topics such as brand guidelines, marketing channels, products, and household information.

Index files tell Claude where to find the relevant detail. A business_profile.md file points to the company overview, course information, partnerships, and other focused files without loading all of them at once.

Her global claude.md routes business requests to the business profile and personal requests to the personal profile, allowing Claude to load the smaller set of context required for the task.

Add context after real work

Teresa grows the library at the end of useful sessions by asking:

Claude, what'd you learn today that we should document?

Claude then proposes updates to the relevant context files. A short request such as "Claude blog post review, gimme feedback" can draw on her writing guide, audience profile, and product information.

What makes this work

The task manager, research digest, and context library all use local Markdown files with enough structure for scripts and Claude to work with them. Teresa can inspect, search, move, and back up those files without depending on one off-the-shelf interface.

Set Up an Automated Search Script Build a Python research searcher driven by a YAML topic file. Search arXiv every day and Google Scholar once a week, deduplicate results by stable paper ID or URL, and write only new results to a dated Markdown digest with title, authors, abstract, source, and link. Include setup and cron examples.

Create an Automated Summarization Script Build a nightly Python script that finds unprocessed PDFs under research/*/sources, invokes a Claude Code agent with my paper summary skill, saves each Markdown summary under the matching notes folder, and writes a manifest so successful files are not processed twice. Report failures without marking them complete.

Engineer a High-Quality Summarization Prompt Analyze this academic paper for critical review. Return: 1. Core claim, 2. Study design and sample, 3. Measures and comparison groups, 4. Main findings with the reported numbers and effect sizes, 5. Limitations and threats to validity, 6. Relevance to [topic]. If the paper does not report an item, say so. Distinguish the authors' claims from your assessment.

Start with one set of information you want to own, such as tasks or research notes. Define a small file format, add one command that produces a useful daily view, and split context into indexed files only when unrelated information begins appearing in the same prompt.

Build a research digest that searches arXiv daily and Google Scholar weekly, filters results against your own topics, and turns selected PDFs into rigorous summaries of methods, findings, effect sizes, and limitations.